Migration model training method, electronic device, and storage medium

Abstract

The application relates to the technical field of deep learning, and provides a transfer model training method, an electronic device and a storage medium. The method obtains a prediction result of an input image corresponding to a transfer scene through a meta model, pre-processes the input image based on a check result of the prediction result of a user, and thus obtains a training sample; the meta model is trained by using a historical sample corresponding to the meta model and the training sample, and a historical gradient corresponding to the historical sample in a training process and a transfer gradient corresponding to the training sample are obtained; a mixed gradient is determined by the similarity between the historical gradient and the transfer gradient, the model parameters of the meta model are updated based on the mixed gradient, and thus a transfer model corresponding to the transfer scene is obtained based on online training of the meta model. The above method can reduce the scene transfer cost of the model and improve the transfer effect.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and more specifically, to a transfer model training method, electronic device, and storage medium. Background Technology

[0002] Deep learning can fit the mapping relationship between input and output by building and training models. In practical applications, due to the limited sample data, online transfer learning techniques are usually used to transfer the capabilities of general models corresponding to general scenarios to new transfer scenarios.

[0003] However, due to the scarcity of samples in the transfer scenarios and the significant differences between the transfer scenarios and the general scenarios, online training of the model may result in the model being effective only for the general scenarios that have already been trained on, thus losing its ability to generalize to the transfer scenarios. Summary of the Invention

[0004] In view of the above, it is necessary to propose a transfer model training method, electronic device and storage medium that can solve the technical problem of insufficient generalization ability of online trained models due to the scarcity of samples in transfer scenarios and the large differences between transfer scenarios and general scenarios.

[0005] An embodiment of this application provides a transfer model training method, the method comprising: preprocessing the input image based on the prediction result of the meta-model on the input image and the user's verification result of the prediction result to obtain training samples, wherein the input image includes an image corresponding to a transfer scene; training the meta-model using historical samples corresponding to the meta-model and the training samples to obtain historical gradients corresponding to the historical samples and transfer gradients corresponding to the training samples; determining a mixed gradient of the historical gradient and the transfer gradient based on the similarity between the historical gradient and the transfer gradient; updating the model parameters of the meta-model based on the mixed gradient; and training the transfer model corresponding to the transfer scene based on the updated model parameters.

[0006] In one embodiment, training the meta-model for each cycle includes: sequentially sampling training samples from the training samples to obtain each batch of training samples, and sequentially sampling historical samples from the historical samples to obtain each batch of historical samples, until all training samples have been sampled, thereby completing the update of the meta-model for the current cycle.

[0007] In one embodiment, the method further includes determining the category of the input image based on the prediction result and the verification result, including: if the verification result indicates that the prediction result corresponds to the target object, determining the category of the input image as a correct sample; if the verification result indicates that the prediction result did not predict the target object, determining the category of the input image as a missed sample; if the verification result indicates that the prediction result incorrectly predicted other objects as the target object, determining the category of the input image as a false positive sample.

[0008] In one embodiment, the preprocessing includes: if the category of the input image is a correct sample, using the meta-model to predict the bounding boxes of the correct sample that are below a preset confidence threshold, selecting the bounding boxes from the bounding boxes below the preset confidence threshold whose intersection-union ratio with the user-labeled bounding boxes is greater than a preset intersection-union ratio threshold, and using the bounding box with the highest confidence among the selected bounding boxes as the labeling information of the correct sample to obtain the training sample.

[0009] In one embodiment, the preprocessing further includes: if the category of the input image is a missed sample, using the user-defined bounding box corresponding to the missed sample as the annotation information of the missed sample to obtain the training sample.

[0010] An embodiment of this application provides a transfer model training apparatus, the apparatus comprising: an input module, configured to preprocess the input image based on the prediction result of the meta-model on the input image and the user's verification result of the prediction result to obtain training samples, wherein the input image includes an image corresponding to a transfer scene; a gradient determination module, configured to train the meta-model using historical samples corresponding to the meta-model and the training samples to obtain historical gradients corresponding to the historical samples and transfer gradients corresponding to the training samples; and a parameter update module, configured to determine a mixed gradient of the historical gradient and the transfer gradient based on the similarity between the historical gradient and the transfer gradient, update the model parameters of the meta-model based on the mixed gradient, and train the transfer model corresponding to the transfer scene based on the updated model parameters.

[0011] An embodiment of this application provides an electronic device, which includes a processor and a memory, wherein the processor is used to implement the transfer model training method when executing a computer program stored in the memory.

[0012] Embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the transfer model training method.

[0013] In summary, the transfer model training method described in this application obtains the prediction results of the input image corresponding to the transfer scenario through a meta-model. Based on the user's verification of the prediction results, the input image is preprocessed to obtain training samples. The meta-model is trained using historical samples and training samples to obtain the historical gradients corresponding to the historical samples and the transfer gradients corresponding to the training samples. A mixed gradient is determined based on the similarity between the historical gradient and the transfer gradient. The model parameters of the meta-model are updated based on the mixed gradient, thereby achieving online training of the meta-model to obtain the transfer model corresponding to the transfer scenario. This method can reduce forgetting and bias phenomena during model transfer, thereby reducing the amount of data required to transfer the model from a general scenario to the transfer scenario, thus reducing the cost of scenario transfer and improving the transfer effect. Attached Figure Description

[0014] Figure 1 This is a structural diagram of an electronic device provided in an embodiment of this application.

[0015] Figure 2 This is a flowchart of a transfer model training method provided in an embodiment of this application.

[0016] Figure 3 This is a flowchart of a method for updating a historical sample database according to an embodiment of this application.

[0017] Figure 4 This is a schematic diagram of the process of training a meta-model using historical samples and training samples corresponding to the meta-model, provided in an embodiment of this application.

[0018] Figure 5 This is a flowchart of a method for determining a hybrid gradient provided in an embodiment of this application.

[0019] Figure 6 This is a flowchart illustrating a transfer model training method provided in an embodiment of this application.

[0020] Figure 7 This is a structural diagram of a transfer model training device provided in an embodiment of this application. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing an embodiment in one instance only and is not intended to be limiting of the application.

[0023] It should be noted that in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and drawings of this application are used to distinguish similar objects, not to describe a specific order or sequence.

[0024] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner. Unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0025] In one embodiment, deep learning can fit the mapping relationship between input and output by building and training models, and is currently widely used in various artificial intelligence tasks, such as object detection and text classification.

[0026] Specifically, the process of deep learning methods typically involves building a model network structure, training the model using a large amount of data, and fitting the model's parameters to a mapping relationship between the input and output data in the current scene. For example, in an object detection task, the input data could be an image containing a "person," and the object detection task could be detecting "people" in the image. In this case, the mapping relationship that the model needs to fit is from "an image containing a person" to "the location and confidence level of the specific person in the image."

[0027] With the same model structure and a limited amount of sample data, the performance of the trained model gradually increases with the increase in the amount of sample data. In practical applications, due to the limited sample data, online transfer learning techniques are typically used to transfer the capabilities of a general model corresponding to a general scenario to a new transfer scenario. Specifically, transfer learning techniques aim to use knowledge and capabilities from a general scenario to guide a model for a specific scenario (a new transfer scenario), enabling the model to improve its generalization performance on the same limited number of samples.

[0028] However, due to the scarcity of samples in the transfer scenarios and the significant differences between the transfer scenarios and the general scenarios, online training of the model may result in the model being effective only for the general scenarios that have already been trained on, thus losing its ability to generalize to the transfer scenarios.

[0029] For example, model transfer can be achieved based on ladder side tuning (LST) technology. LST technology can add a side network to a general model to correct the model results. By using data from specific scenarios to train the corresponding ladder side network, the goal is to reduce model overfitting and enhance the model's generalization performance in specific scenarios while retaining the capabilities of the original general model.

[0030] However, the training samples collected during model training for a specific scenario may differ from the samples in the corresponding real-world application scenario. To adapt the model to real-world application scenarios, it is necessary to further transfer the model's tasks to these scenarios. Specifically, samples from real-world application scenarios can be added to the ladder-side network for further training, allowing the network to refit to the real-world application scenario. This training process uses fewer training samples compared to other training processes and can be called an online training process.

[0031] However, during online training, since the number of available samples in real-world application scenarios is often very small, and the task definition in real-world application scenarios may not be consistent with the general scenario, bias can easily occur during training, causing the model to be effective only for the scenarios that have already been trained on and lose its generalization ability.

[0032] The above phenomenon is more likely to occur when the level of detail in actual application scenarios leads to a mismatch between capabilities in general scenarios and those in specific scenarios. This mismatch can be roughly divided into the following two situations:

[0033] (1) The tasks in general scenarios have become more refined in actual application scenarios.

[0034] For example, a general-purpose model might be trained to detect cigarettes, but a real-world application requires detecting cigarettes that are being smoked, while discarded cigarettes don't need to be detected. Clearly, there's a conflict between the capabilities of the model for the real-world application and the capabilities of the model for the general-purpose application.

[0035] Specifically, a cigarette detection model trained in a general scenario might use training data containing images of cigarettes in various forms and perspectives, including those being smoked, stationary, and discarded. However, in real-world applications where the model is specifically required to detect only cigarettes being smoked, the way the training data describes cigarettes in a "non-smoking state" (especially discarded cigarettes) can become problematic.

[0036] For example, images used to describe non-detectable targets (i.e., cigarettes that don't need to be detected, such as discarded cigarettes) are primarily taken from an overhead perspective. This means that during training, the model may learn a "bias" or "pattern"—that cigarettes seen from an overhead perspective do not need to be detected. This bias may not be a major problem when the training data is sufficient and evenly distributed, but in real-world applications, if the model encounters cigarettes in similar non-smoking states from other perspectives (such as side views, upward views, etc.) (e.g., carelessly placed cigarette butts), it may incorrectly detect these cigarettes as targets that need to be detected due to a lack of negative samples (i.e., targets that do not need to be detected) from the corresponding perspective.

[0037] Due to the aforementioned perspective bias, a large number of false detections may occur when the model is applied to real-world scenarios. Specifically, the model may incorrectly identify cigarettes that do not actually need to be detected (such as cigarette butts from other perspectives) as cigarettes in the smoking process that need to be detected.

[0038] (2) There is a fusion between the tasks in general scenarios and the tasks in actual application scenarios.

[0039] For example, in a general scenario, the model is trained to detect objects of different colors. However, in a real-world application, the task requires both sphere detection and red light detection. In this case, there is a fusion between the tasks in the general scenario and the tasks in the real-world application; for instance, both tasks in the general scenario and the tasks in the real-world application require color detection.

[0040] Models can become confused when different tasks have different requirements for the same input feature (e.g., color). For example, the model's feature requirements for multiple tasks may conflict. Specifically, the sphere detection task itself does not care about the color of the sphere; it mainly focuses on features such as the shape and possible texture of the object. Red light detection, on the other hand, relies heavily on color information, especially the specific color red. Since the model has already learned how to distinguish objects based on color during training in previous general scenarios, it may tend to rely on color as a feature when faced with the red light detection task.

[0041] Furthermore, because the model may have already tightly linked color and object detection during training in general scenarios, it may incorrectly use color information as a crucial basis for judging the detection target when performing multi-task tasks. This could lead to a green light (the color of which is similar to that of a red light but is not a detection target) being misdetected as a red light, because the model may focus too much on color and ignore other more critical features (such as shape, position, etc.).

[0042] On the other hand, since sphere detection tasks do not require color information, the model may gradually ignore the importance of color as a feature for sphere detection during the learning process. This could lead to situations where, even if a green ball (whose color is similar to that of a green light) is visually very obvious, it may be missed because the model no longer prioritizes color information and relies excessively on other potentially less significant features (such as shape).

[0043] Based on the above analysis, the aforementioned bias problem can cause the ladder network to produce incorrect correction directions and strengths after the model has been fully fitted, reducing the model's generalization ability. It also damages the model's original task capabilities, causing forgetting phenomena, and thus the model's ability after online training decreases instead of increasing.

[0044] To address the aforementioned issues, this application provides a transfer model training method. The method involves obtaining prediction results for the input image corresponding to the transfer scenario through a meta-model, preprocessing the input image based on user verification of the prediction results to obtain training samples, training the meta-model using historical samples and training samples, and obtaining historical gradients for historical samples and transfer gradients for training samples. A mixed gradient is determined based on the similarity between the historical and transfer gradients, and the model parameters of the meta-model are updated based on the mixed gradient. This achieves online training of the meta-model to obtain the transfer model corresponding to the transfer scenario. This method reduces forgetting and bias phenomena during model transfer, thereby reducing the amount of data required to transfer the model from a general scenario to the transfer scenario, thus lowering the scenario transfer cost and improving the transfer effect.

[0045] Figure 1 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 10 can be used in electronic devices such as computers, servers, mobile phones, tablets, and laptops. This embodiment of the application does not impose any restrictions on the specific type of electronic device.

[0046] like Figure 1 As shown, the electronic device 10 may include a communication module 101, a memory 102, a processor 103, an input / output (I / O) interface 104, and a bus 105. The processor 103 is coupled to the communication module 101, the memory 102, and the I / O interface 104 via the bus 105.

[0047] Communication module 101 may include a wired communication module and / or a wireless communication module. The wired communication module may provide one or more wired communication solutions such as Universal Serial Bus (USB) and Controller Area Network (CAN). The wireless communication module may provide one or more wireless communication solutions such as Wireless Fidelity (Wi-Fi), Bluetooth (BT), mobile communication networks, Frequency Modulation (FM), Near Field Communication (NFC), and Infrared (IR).

[0048] Memory 102 may include one or more random access memory (RAM) and one or more non-volatile memory (NVM). The RAM can be directly read and written by the processor 103, and can be used to store executable programs (e.g., machine instructions) of the operating system or other running programs, as well as user and application data. The RAM may include static random-access memory (SRAM), dynamic random-access memory (DRAM), synchronous dynamic random-access memory (SDRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.

[0049] Non-volatile memory can also store executable programs and user and application data, and can be pre-loaded into random access memory for direct reading and writing by the processor 103. Non-volatile memory can include disk storage devices and flash memory.

[0050] Memory 102 is used to store one or more computer programs. The one or more computer programs are configured to be executed by processor 103. The one or more computer programs include multiple instructions that, when executed by processor 103, enable a transfer model training method to be executed on electronic device 10.

[0051] In other embodiments, the electronic device 10 further includes an external memory interface for connecting to an external memory to expand the storage capacity of the electronic device 10.

[0052] Processor 103 may include one or more processing units, such as application processor (AP), modem processor, graphics processing unit (GPU), image signal processor (ISP), controller, video codec, digital signal processor (DSP), baseband processor, and / or neural network processing unit (NPU). These different processing units may be independent devices or integrated into one or more processors.

[0053] The processor 103 provides computational and control capabilities, for example, the processor 103 is used to execute computer programs stored in memory 102 to implement the transfer model training method described above.

[0054] I / O interface 104 is used to provide a channel for user input or output. For example, I / O interface 104 can be used to connect various input and output devices, such as mouse, keyboard, touch device, display screen, etc., so that users can enter information or visualize information.

[0055] Bus 105 is used at least to provide a channel for communication between communication modules 101, memory 102, processor 103, and I / O interface 104 in electronic device 10.

[0056] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the electronic device 10. In other embodiments of this application, the electronic device 10 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0057] Figure 2 This is a flowchart of a transfer model training method provided in an embodiment of this application. The transfer model training method is applied to electronic devices, for example... Figure 1 The electronic device 10 in the process includes the following steps. Depending on different needs, the order of the steps in the flowchart can be changed, and some steps can be omitted.

[0058] Step S21: Based on the prediction results of the meta-model on the input image and the user's verification results of the prediction results, the input image is preprocessed to obtain training samples.

[0059] In one embodiment, the model transfer process may include iterative training of a model corresponding to a general scenario over multiple cycles. During each cycle of model training, the model parameters can be tuned to gradually transfer the model to the actual application scenario and ensure high performance of the resulting transferred model. The transfer model training method provided in this application embodiment can be considered as a method for online training of the transfer model by tuning model parameters in a certain cycle (e.g., including multiple training sessions) or a single session.

[0060] In one embodiment, the meta-model can be iteratively trained over multiple cycles. In the first cycle of this iterative training, the meta-model includes a pre-trained model corresponding to a historical scenario (e.g., a general scenario). In the (N+1)th cycle of the iterative training, the meta-model includes the model obtained after the Nth cycle of training, where N represents a natural number greater than or equal to 1. This allows for iterative updates to the meta-model, gradually migrating it to real-world application scenarios (e.g., scenario migration).

[0061] For example, the task corresponding to a historical scenario could be object detection (e.g., a cigarette), and the pre-trained model for the historical scenario could be a first object detection model. The task corresponding to a transfer scenario could be object detection of a target object in a certain environment (e.g., the process of smoking).

[0062] In one embodiment, during the transfer training of the meta-model, sample images corresponding to the transfer scenario can be used, such as user-input images. Since input images typically do not contain label information (e.g., predicted bounding boxes and predicted object categories), and manually labeling input images is time-consuming and labor-intensive, a meta-model can be used to assist in labeling input images, thereby improving the efficiency of obtaining training samples.

[0063] Specifically, a meta-model can be used to predict the input image and obtain the prediction result. In one example, the prediction result may include the predicted bounding box corresponding to the target object and the predicted object category. The predicted bounding boxes may include bounding boxes with a confidence level higher than a preset confidence threshold. The confidence threshold can be set according to actual needs, for example, 0.9; this application does not impose a specific limitation on this. The predicted bounding boxes may also include multiple bounding boxes. For example, the multiple bounding boxes may include multiple (e.g., 3) bounding boxes with different shapes and sizes, and the confidence level of each of the multiple bounding boxes is greater than the confidence threshold.

[0064] Because the meta-model may have low prediction accuracy for the input image corresponding to the migration scenario, and the prediction results may include errors, the meta-model cannot determine whether the prediction result is accurate. By receiving user verification results of the prediction results, the input image can be preprocessed to obtain the calibrated label for the input image.

[0065] For example, user verification of the prediction results may include whether the prediction is accurate, and preprocessing the input image may include updating the predicted bounding boxes and the predicted object categories. It is understandable that since the meta-model has a certain prediction accuracy for the input image, preprocessing is not necessary for input images with accurate predictions.

[0066] Through the above embodiments, the efficiency of obtaining labels for input images can be improved based on meta-models, and the accuracy of sample labels can be ensured to meet the requirements based on the prediction results of meta-models and the verification results of the prediction results received by users.

[0067] In one embodiment, the system may also receive supplementary annotations from the user on the input image. These supplementary annotations may include user-annotated bounding boxes obtained by updating the predicted bounding boxes based on the location of the target object.

[0068] In one example, the predicted bounding boxes obtained from the metamodel can be displayed to the user in the input image, and user-annotated boxes can be obtained in response to the user's update operations on the position and / or shape of the predicted bounding boxes.

[0069] In one embodiment, the category of the predicted input image obtained by the meta-model can be determined based on the prediction and verification results, further determining the category of the corresponding input image. This allows for a determination of whether the predicted input image needs to be updated based on the user's supplementary annotations, depending on the category of the input image.

[0070] In one embodiment, determining the category of an input image based on the prediction result and the verification result may include: if the verification result indicates that the prediction result corresponds to the target object, determining the category of the input image as a correct sample; if the verification result indicates that the prediction result did not predict the target object, determining the category of the input image as a missed sample; if the verification result indicates that the prediction result incorrectly predicted other objects as the target object, determining the category of the input image as a false detection sample.

[0071] Specifically, the verification result may include whether the prediction result is accurate, including whether the predicted bounding box and the predicted object category are accurate. The following embodiment will use the classification of the input image based on the accuracy of the predicted bounding box as an example. The method for classifying the input image based on the accuracy of the predicted object category can be referred to in the following embodiment.

[0072] In one embodiment, if the verification result indicates that both the predicted bounding box and the predicted object category in the prediction result are accurate, it can be determined that the verification result corresponds to the target object, and such an input image can be identified as a correct sample. A correct sample can represent a sample that is both correctly detected by the meta-model (i.e., the meta-model believes that the target object exists in the region) and that these detections actually correspond to the real target (i.e., the user-defined bounding box also marks the region as the target object). The accuracy of the predicted bounding box can be determined based on the overlap between the user-defined bounding box and the predicted bounding box. For example, the intersection-union ratio (IUR) between the user-defined bounding box and the predicted bounding box can be used as the overlap. If the IUR is greater than or equal to a preset value (e.g., 0.3), the predicted bounding box can be determined to be relatively accurate, and such a bounding box can be identified as accurate.

[0073] In one embodiment, if the verification result indicates that the predicted bounding box in the prediction result is inaccurate, it can be determined that the prediction result did not predict the target object, and such an input image is identified as a missed sample. A missed sample can represent a sample in which the target object actually exists in the input image, but the meta-model failed to detect it correctly. That is, the user-labeled bounding box marks a region as the target object, but the meta-model does not provide any prediction for that region, or the meta-model considers the confidence level of that region too low and filters it out. For example, if the intersection-over-union ratio (IoU) between the user-labeled bounding box and the predicted bounding box is less than a preset value (e.g., 0.3), it can be determined that the predicted bounding box is inaccurate, and the prediction result did not predict the target object.

[0074] In one embodiment, if the verification result indicates that the bounding box in the prediction result is not the target object, it can be determined that the prediction result incorrectly predicted other objects as the target object, and the input image is identified as a false detection sample. A false detection sample can represent a sample that is incorrectly detected as a target object by the meta-model (i.e., the meta-model believes that a target object exists in the region), but in reality, the input image or region does not contain a true target object. In object detection tasks, such false detections typically occur when the model generates excessively high confidence in non-target regions such as image background or noise, thus incorrectly predicting the existence of the target object.

[0075] In one embodiment, different preprocessing strategies can be adopted for different types of samples to optimize training data and improve model performance. For example, false positive samples may not contain the real target object or the predicted bounding box corresponding to the target object, so they can be directly used as training samples (e.g., negative samples that do not contain the target object). However, for correct samples and false negative samples that contain the predicted bounding box of the target object, more complex preprocessing is required, such as correcting the predicted bounding box based on the prediction results of the meta-model and the user-annotated box.

[0076] In one embodiment, the preprocessing includes: if the category of the input image is a correct sample, using a meta-model to predict the bounding boxes of the correct samples that are below a preset confidence threshold, selecting bounding boxes from the bounding boxes below the preset confidence threshold whose intersection-union ratio (IU) with the user-labeled bounding boxes is greater than a preset IU threshold, and using the bounding box with the highest confidence among the selected bounding boxes as the labeling information of the correct sample to obtain the training sample.

[0077] In one example, the predicted bounding boxes in the correct sample are bounding boxes with high confidence determined by the meta-model. However, due to the potentially poor prediction accuracy of the meta-model, the actual overlap between the predicted bounding boxes and the accurate user-labeled boxes may not be very high (e.g., intersection-union ratio less than 0.5). The meta-model can be used to predict bounding boxes below a preset confidence threshold for the correct sample. While these bounding boxes are below the meta-model's confidence threshold, they may actually have a high overlap with the accurate user-labeled boxes (e.g., intersection-union ratio greater than 0.5). Bounding boxes with an intersection-union ratio greater than a preset threshold (e.g., 0.5) with the user-labeled boxes can be selected from the bounding boxes below the preset confidence threshold. The bounding box with the highest confidence among the selected bounding boxes is then used as the labeling information for the correct sample.

[0078] Through the above embodiments, annotation information can be optimized based on the feedback of the model's prediction results. Specifically, the meta-model's prediction results reflect its understanding of the current data. If the meta-model can detect a target object at a certain location (even if it is below a preset confidence threshold), then this location is likely to be a region where the target object actually exists. By selecting the prediction box with the highest confidence level and an intersection-union ratio (IU) above the IU threshold as the correction, this feedback from the model is actually being used to optimize the annotation information.

[0079] In addition to the first type of training samples obtained by optimizing the annotation information of correct samples based on model feedback, the user-defined bounding boxes corresponding to the correct samples can also be used as the annotation information (e.g., bounding boxes) of the correct samples to obtain the second type of training samples. Using both the first and second type of training samples as training samples can be regarded as a data augmentation strategy.

[0080] By incorporating model predictions to refine bounding boxes, more training samples can be provided for model training. These samples may differ from user-defined bounding boxes in the position and shape of the target object, but they are closer to the model's current understanding. This makes the annotation information more adaptable to the characteristics of the current model, helping the model learn more robust feature representations, thereby improving the model's training efficiency and performance.

[0081] In one embodiment, the preprocessing further includes: if the category of the input image is a missed sample, using the user-defined bounding box corresponding to the missed sample as the annotation information of the missed sample to obtain training samples.

[0082] In one example, for missed samples where the meta-model failed to detect the target object at the corresponding location, the user-defined bounding box can be used directly as the corresponding annotation information to obtain training samples.

[0083] The above embodiments ensure the integrity and accuracy of the annotation information of the training samples, while helping the model learn how to detect target objects that were previously missed.

[0084] Step S22: Train the meta-model using the historical samples and training samples corresponding to the meta-model to obtain the historical gradient corresponding to the historical samples and the transfer gradient corresponding to the training samples.

[0085] In one embodiment, the meta-model can be trained using a gradient-based network training method (hereinafter referred to as the meta-model training method), and the model parameters of the meta-model can be updated using a gradient descent algorithm (or its variants, such as the Adam algorithm, RMSprop algorithm, etc.).

[0086] In one embodiment, the meta-model training method provided in this application is broadly similar to the training process of a normal deep learning model. However, during each iteration update of the meta-model, the meta-model training method uses not only the gradients corresponding to the training samples used in the normal deep learning training process, but also the historical gradients of historical samples representing the original knowledge of the meta-model. Furthermore, during each iteration, the learning strategy is adjusted based on the difference between these historical gradients and the gradients of new samples (e.g., the transfer gradients corresponding to the training samples). For example... Figure 4The diagram shown is a flowchart illustrating the process of training a meta-model using historical samples and training samples corresponding to the meta-model, according to another embodiment of this application.

[0087] In one embodiment, based on the above meta-model training method, historical samples and training samples can be sampled simultaneously during meta-model training. Each batch of samples obtained from the sampling is used to perform each iteration of the meta-model in each cycle. Since the number of training samples may be much smaller than the number of historical samples, the current training cycle can be considered complete after sampling all batches of training samples, and the next cycle's training can begin with new training samples.

[0088] In one embodiment, training the meta-model for each cycle includes: sequentially sampling training samples from the training samples to obtain each batch of training samples, and sequentially sampling historical samples from the historical samples to obtain each batch of historical samples, until all training samples have been sampled, thus completing the update of the meta-model for the current cycle.

[0089] In one example, during the sampling of historical samples, the number of samples in each batch is the same as the number of samples in the training samples. Furthermore, the sampling process for historical samples is independent and is not reset with the sampling process of the training samples or with periodic updates, thus ensuring that the historical samples used in different periods are different.

[0090] In another embodiment, if the number of training samples is very small (e.g., less than the preset number of samples per batch), the number of samples may only need to be sampled once per period. In this case, the transfer model training method provided in this application embodiment may also include a method of training the meta-model once.

[0091] Through the above embodiments, efficient and comprehensive updates to the meta-model can be achieved by simultaneously utilizing historical and training samples during the meta-model training process. By sequentially sampling training and historical samples and ensuring that the sampling quantities of both match in each cycle, not only can rich historical data be fully utilized, but the timeliness and effectiveness of new training samples can also be ensured, thereby accelerating the convergence speed of the meta-model and improving its generalization ability and accuracy.

[0092] In one embodiment, the method for determining the historical gradient corresponding to a historical sample is similar to the method for determining the transfer gradient corresponding to a training sample. In the following embodiments, the determination of the historical gradient corresponding to a historical sample will be used as an example.

[0093] In one embodiment, historical samples can be input into the meta-model, and the prediction results of these historical samples can be obtained through the model's forward propagation process. The difference between the prediction results and the labels of the historical samples can be quantified using a loss function (e.g., mean squared error, cross-entropy loss, etc.) to obtain the loss value corresponding to the historical samples. Then, the gradients of these loss functions with respect to the model parameters can be calculated using the backpropagation algorithm. Specifically, the backpropagation algorithm can start from the loss function and calculate the gradient of the loss function with respect to the parameters (e.g., connection weights) of each layer of the model, layer by layer, until it reaches the input layer. During this process, the chain rule is used to propagate the gradient of the loss function with respect to the output layer back to the input layer, thereby obtaining the historical gradients of the model parameters.

[0094] Through the above embodiments, the gradient information of the meta-model during the training process can be obtained, which facilitates updating the model parameters based on the gradient information and optimizing the model.

[0095] In one embodiment, the historical sample library, consisting of historical samples corresponding to the meta-model, is updated with each training cycle. The simplest update method is to use the training samples corresponding to the Nth cycle as the updated historical samples corresponding to the N+1th cycle, thereby continuously expanding the sample richness of the historical sample library. Here, N represents an integer greater than or equal to 1. When N=1, the historical samples corresponding to the meta-model represent the training samples corresponding to the historical scene, and the historical samples are labeled training samples.

[0096] In one embodiment, after completing the training of the meta-model for each cycle, the historical sample library composed of historical samples can be updated according to the historical gradients corresponding to historical samples and the transfer gradients corresponding to training samples, and the samples in the updated historical sample library (e.g., updated historical samples) can be used to train the meta-model for the next cycle.

[0097] In one embodiment, when updating the historical sample library based on historical gradients and migration gradients, historical samples in the historical sample library that conflict with the migration scenario can be identified based on the distance between the historical gradients and migration gradients, and these historical samples can be removed from the historical sample library. In another example, the method for updating the historical sample library can also refer to... Figure 3 The flowchart shown is shown.

[0098] Through the above embodiments, training samples in the training cycle can be continuously transformed into historical samples, and the historical sample library can be updated based on the gradient information of historical samples and current training samples, effectively improving the sample richness and adaptability of the meta-model. Not only can the historical sample library be expanded, but also, through sample updates guided by gradient information, the meta-model can better capture changes in data distribution, thereby adapting more quickly and making accurate predictions when facing new tasks in transfer scenarios.

[0099] Step S23: Based on the similarity between the historical gradient and the transfer gradient, determine the mixed gradient of the historical gradient and the transfer gradient, update the model parameters of the meta-model based on the mixed gradient, and train the transfer model corresponding to the transfer scenario based on the updated model parameters.

[0100] In one embodiment, historical gradients and transfer gradients can be blended based on their similarity. The blended gradients are used to dynamically adjust the model parameters of the meta-model, and then the updated meta-model parameters are used to train the transfer model for the transfer scenario, thereby achieving accurate parameter optimization and rapid adaptation to the transfer scenario. The similarity can be cosine similarity, and this embodiment does not impose specific limitations on it. In one example, the method for determining the blended gradients can also refer to... Figure 5 The flowchart shown is shown.

[0101] In one embodiment, dynamically adjusting the model parameters of the metamodel using mixed gradients can be accomplished by an optimizer (such as SGD, Adam, etc.). The optimizer adjusts the values ​​of the model parameters based on gradient information (e.g., mixed gradients) and the learning rate to minimize the loss function. For example, the optimizer can use strategies such as momentum, learning rate decay, and gradient descent algorithms to accelerate the convergence process and avoid getting trapped in local optima.

[0102] The transfer model training method provided in this application can transfer the model to a new scenario (e.g., a new scenario) during online learning, and retain the model's general capabilities as much as possible while ensuring that the historical and new scenarios do not conflict. It leverages the knowledge of large-scale models in general domains to improve the model's application performance in new scenarios with fewer samples, reducing the amount of sample data required for training the model in new scenarios.

[0103] In one embodiment, for example Figure 3 As shown, the method for updating the historical sample database may include the following process:

[0104] Step S31: Use the training sample corresponding to the Nth period as the updated historical sample corresponding to the N+1th period, and determine the average gradient of the training sample corresponding to the Nth period.

[0105] In one embodiment, N represents a natural number greater than or equal to 1. By using the training sample corresponding to the Nth period as the updated historical sample corresponding to the N+1th period, the sample richness of the historical sample library can be continuously expanded.

[0106] In one embodiment, the training samples corresponding to the Nth period may include multiple training samples. The average gradient of the training samples corresponding to the Nth period can be determined based on the ratio of the sum of all training samples corresponding to the Nth period to the total number of training samples corresponding to the Nth period.

[0107] Step S32: Determine the difference between the historical gradient and the average gradient of each historical sample in the historical samples corresponding to the Nth period, and delete the preset number of historical samples with the largest difference from the historical sample library corresponding to the Nth period.

[0108] In one embodiment, the historical samples corresponding to the Nth period may include multiple historical samples, and the degree of difference between the historical gradient and the average gradient of each historical sample can be determined. For example, the inverse of the cosine distance between the historical gradient and the average gradient of each historical sample can be used as the degree of difference. The greater the degree of difference, the greater the conflict between the corresponding historical sample and the migration scenario.

[0109] In one embodiment, a preset number of samples with the highest degree of difference can be deleted from the historical sample library corresponding to the Nth period, based on the degree of difference. The preset number can be set according to actual needs. In one example, historical samples with a degree of difference greater than a preset degree of difference threshold (e.g., 0) can also be selected as the historical samples to be deleted, or a preset number of historical samples with a degree of difference greater than the preset degree of difference threshold and the highest degree of difference can be selected as the historical samples to be deleted.

[0110] Step S33: Construct an updated historical sample library based on the updated historical samples and the historical sample library corresponding to the Nth period after deletion.

[0111] In one embodiment, the latest training samples can be included in the historical sample library as updated historical samples, and then the historical sample library after removing historical samples with large differences in average gradient from the training samples can be merged to obtain the updated historical sample library.

[0112] In one embodiment, the formula corresponding to the method for updating the historical sample database includes:

[0113] S new =Strain +S old -S del ;

[0114] S del ={x|Top(-CosSim(g,g x ),K)}∩{x|-CosSim(g,g x )>0};

[0115] Among them, S new S represents the historical sample database corresponding to the (N+1)th period (e.g., the updated historical sample database). train S represents the training sample (e.g., an updated historical sample) corresponding to the Nth period. old S represents the historical sample library corresponding to the Nth period. del Let x represent the set of historical samples to be deleted (e.g., a preset number of historical samples with the highest dissimilarity), and g represent the average gradient of the training samples corresponding to the Nth period. x This represents the historical gradient of historical sample x corresponding to the Nth period, K represents the preset number, CosSim() represents the cosine similarity function, and -CosSim(g,g) x ) represents the historical gradient g x The degree of difference between the average gradient g and the mean gradient g, where N represents a natural number greater than or equal to 1.

[0116] Through the above embodiments, the latest training samples can be incorporated into the historical sample library to enrich its diversity and achieve dynamic updates to the historical sample library. Furthermore, by calculating and comparing the difference in average gradient between historical samples and training samples in the current training cycle, historical samples that most strongly conflict with the current learning trend are accurately removed. This not only ensures the timeliness and relevance of the historical sample library but also improves the efficiency and accuracy of model training by reducing interference from irrelevant or conflicting samples, thereby enhancing the model's adaptability and generalization performance when transferred to new scenarios.

[0117] In one embodiment, for example Figure 5 As shown, the method for determining the hybrid gradient may include the following process:

[0118] Step S51: Determine the weights corresponding to the historical gradient and the weights corresponding to the migration gradient based on the similarity.

[0119] In one embodiment, the weights of the historical gradient and the transfer gradient can be dynamically adjusted based on their similarity. For example, a high similarity indicates a lower degree of conflict between the historical sample and the transfer scenario, allowing the historical gradient to receive a larger weight to facilitate the model's effective use of past experience. Conversely, a low similarity indicates a higher degree of conflict between the historical sample and the transfer scenario, allowing the transfer gradient weight to be increased, thereby enabling the model to adapt to changes in the new scenario more quickly.

[0120] In one embodiment, the weights corresponding to the migration gradient can be inversely proportional to the similarity, while the weights corresponding to the historical gradient can be directly proportional to the similarity.

[0121] Through the above embodiments, an intelligent balance can be achieved between the historical knowledge corresponding to historical samples and the new knowledge corresponding to training samples during the transfer learning process, thereby enhancing the flexibility and adaptability of the model and ensuring the effectiveness of model transfer learning.

[0122] Step S52: Determine the mixed gradient based on the weighted sum of the historical gradient and the migration gradient.

[0123] In one embodiment, historical gradients and migration gradients can be mixed using a weighted average method, and the mixed gradients can be used to dynamically adjust the model parameters of the meta-model to obtain updated model parameters of the meta-model.

[0124] In one embodiment, the model parameters of the meta-model are dynamically adjusted using the hybrid gradient, and the formula used may include:

[0125] θ * =θ-LR×(α×g his ×Loss jis +(1-α)×g trans ×Loss trans );

[0126]

[0127] Where θ represents the model parameters, θ * This represents the updated model parameters; LR and β represent the preset hyperparameters, and LR represents the preset learning rate; g his Represents the historical gradient, Loss his g represents the loss value corresponding to the historical sample. trans Represents the migration gradient, Loss trans This represents the loss value corresponding to the training sample; CosSim() represents the cosine similarity function.

[0128] In one embodiment, the historical gradient g his With the migration gradient g trans The similarity between CosSim(g)his ,g trans The value range of ) is [-1,1], and can be determined according to CosSim(g his ,g trans This determines whether updates to historical samples should be received. For example, in the extreme case, when CosSim(g his ,g trans When ) = -1, it indicates the historical gradient g his With the migration gradient g trans The gradients are in opposite directions. In this case, it's determined that there's a task divergence between the historical scene and the transfer scene on the historical sample, and the model should be fitted to the transfer scene. When CosSim(g)... his ,g trans When ) = 1, it indicates the historical gradient g his With the migration gradient g trans If the gradient directions are the same, then it is determined that there is no task divergence between the historical scene and the migration scene on the historical sample, and the model can be fitted to the historical scene.

[0129] In one embodiment, the hyperparameter β ranges from [0,1] and can be set based on prior data. The hyperparameter β can be used to control the maximum proportion of historical knowledge that can be forgotten. For example, when β = 0, α equals 0.5, which means that historical knowledge and new knowledge are equally important for the desired transfer scenario. When β = 1, it means that when there is a conflict between the historical scenario and the transfer scenario, the result of the transfer scenario is adopted.

[0130] In one embodiment, α represents the weight corresponding to the historical gradient, (1-α) represents the weight corresponding to the transfer gradient, and the value of α ranges from [0,1]. When α equals 0.5, the updated model parameters can represent the result of feeding historical samples and training samples into the training simultaneously without performing additional gradient mixing operations.

[0131] Through the above embodiments, the weights of historical gradients and transfer gradients can be dynamically adjusted. The similarity (cosine similarity) between historical and transfer gradients, along with preset hyperparameters, can be used to balance the impact of historical and new knowledge on model parameter updates. This gradient fusion strategy intelligently determines when to prioritize experience from historical scenarios and when to prioritize knowledge from new transfer scenarios during model transfer, thereby effectively improving the model's adaptability and generalization ability in transfer tasks while reducing the negative impact of task divergence.

[0132] refer to Figure 6The diagram shows a flowchart of a transfer model training method provided in another embodiment of this application. Specifically, the transfer model training method may include the following steps: (a) Data acquisition and preprocessing: The system can receive images corresponding to a new scene input by the user, input the input images into the current model N (representing an integer greater than or equal to 1) to obtain detection results; based on the results of manual judgment of whether the detection results are correct or not, the input images are preprocessed. For example, only the detection results of manually verified samples are considered reliable, and the manually verified samples are included in the training samples, while the remaining samples are discarded and do not participate in training. (b) Ladder-side network training and dynamic adjustment of training strategy: The system uses the training samples obtained in (a) and the historical samples corresponding to the model to train the model simultaneously, and mixes the two gradients according to the similarity between the gradients corresponding to the training samples and the gradients corresponding to the historical samples; the mixed gradients are used to update the model parameters, and the model N+1 is trained. (c) Automatic selection of training samples: After each model training is completed, the historical samples will undergo one round of iteration. The data completed in the iteration has two sources. The first source is the training set input in the current round. This batch of data is all included in the training set, and its overall average gradient direction is recorded. Another batch of data is the historical sample pool for this round of training. The gradient of each sample in this part of the sample is calculated to be similar to the average gradient of the training samples. The samples are sorted according to the similarity. If the size of the training set is K, the discard set is constructed as the top K samples with the lowest similarity (or the largest difference) and the similarity is less than 0. The historical samples are excluded from the discard set and put into the historical sample dataset, thus completing the automatic selection of training samples.

[0133] The transfer model training method provided in this application embodiment can, on the basis of the transfer effect of the ladder side network, further reduce the incorrect correction of the original model knowledge when transferring to a new scene during the online training process, reduce the forgetting and bias phenomena in the transfer process, and thus reduce the amount of data required to transfer from a general model to a specific scene, thereby reducing the scene transfer cost and improving the transfer effect.

[0134] Figure 7 This is a structural diagram of a transfer model training device provided in an embodiment of this application.

[0135] In some embodiments, the transfer model training apparatus 70 may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the transfer model training apparatus 70 may be stored in the memory of an electronic device and executed by at least one processor to perform (see details). Figure 2 (Description) Functionality of transfer model training.

[0136] In this embodiment, the transfer model training device 70 can be divided into multiple functional modules according to its functions. These functional modules may include: a preprocessing module 701, a gradient determination module 702, and a parameter update module 703. As used in this application, a module refers to a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functional implementation of each module in the transfer model training device 70 can be found in the above description of the transfer model training method, and will not be repeated here.

[0137] The preprocessing module 701 is used to preprocess the input image based on the prediction result of the meta-model and the user's verification result of the prediction result to obtain training samples, wherein the input image includes images corresponding to the migration scene.

[0138] The gradient determination module 702 is used to train the meta-model using the historical samples corresponding to the meta-model and the training samples to obtain the historical gradient corresponding to the historical samples and the transfer gradient corresponding to the training samples.

[0139] The parameter update module 703 is used to determine the mixed gradient of the historical gradient and the migration gradient based on the similarity between the historical gradient and the migration gradient, update the model parameters of the meta-model based on the mixed gradient, and train the migration model corresponding to the migration scenario based on the updated model parameters.

[0140] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, and the method implemented when the program instructions are executed can refer to the methods in the above embodiments of this application.

[0141] The computer-readable storage medium can be the internal memory of the electronic device described in the above embodiments, such as the hard disk or memory of the electronic device. Alternatively, the computer-readable storage medium can be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device.

[0142] In some embodiments, the computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required for at least one function, etc.; and the data storage area may store data created based on the use of the electronic device, etc.

[0143] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0144] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0145] In the embodiments provided in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0146] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0147] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A transfer model training method, characterized in that, The method includes: Based on the prediction results of the meta-model on the input image and the user's verification results of the prediction results, the input image is preprocessed to obtain training samples, wherein the input image includes images corresponding to the migration scene; The meta-model is trained using the historical samples corresponding to the meta-model and the training samples to obtain the historical gradient corresponding to the historical samples and the transfer gradient corresponding to the training samples. Determining a mixed gradient of the historical gradient and the migration gradient based on the similarity between the historical gradient and the migration gradient includes: determining the weights corresponding to the historical gradient and the migration gradient based on the similarity; and determining the mixed gradient based on the weighted sum of the historical gradient and the migration gradient; and The model parameters of the meta-model are updated based on the hybrid gradient, and the transfer model corresponding to the transfer scenario is trained based on the updated model parameters. After completing the training of the meta-model for each cycle, the historical sample library composed of the historical samples is updated according to the historical gradients corresponding to the historical samples and the transfer gradients corresponding to the training samples, and the samples in the updated historical sample library are used to train the meta-model for the next cycle. The step of updating the historical sample library composed of the historical samples based on the historical gradients corresponding to the historical samples and the migration gradients corresponding to the training samples includes: taking the training samples corresponding to the Nth period as the updated historical samples corresponding to the N+1th period, and determining the average gradient of the training samples corresponding to the Nth period, where N represents a natural number greater than or equal to 1; determining the difference between the historical gradient of each historical sample in the historical samples corresponding to the Nth period and the average gradient, and deleting a preset number of historical samples with the largest difference from the historical sample library corresponding to the Nth period; and constructing an updated historical sample library based on the updated historical samples and the historical sample library corresponding to the Nth period after the deletion operation.

2. The transfer model training method according to claim 1, characterized in that, The method further includes: The meta-model is trained iteratively over multiple cycles, wherein: In the first training cycle of the multiple iterative training cycles, the meta-model includes a pre-trained model corresponding to the historical scene. In the training of the (N+1)th cycle of the multiple iterative training cycles, the meta-model includes the model obtained after the training of the Nth cycle, where N represents a natural number greater than or equal to 1.

3. The transfer model training method according to claim 1, characterized in that, The step of determining the weights corresponding to the historical gradient and the migration gradient based on the similarity includes: Let the weights corresponding to the migration gradient be inversely proportional to the similarity; Let the weights corresponding to the historical gradients be proportional to the similarity.

4. The transfer model training method according to claim 1, characterized in that, The prediction result includes the predicted bounding box corresponding to the target object and the predicted object category. The predicted bounding box includes bounding boxes with a confidence level higher than a preset confidence threshold.

5. The transfer model training method according to claim 4, characterized in that, The method further includes: The system receives supplementary annotation results from the user on the input image. The supplementary annotation results include user-annotated bounding boxes obtained by updating the predicted bounding boxes based on the location of the target object.

6. An electronic device, characterized in that, The electronic device includes: One or more processors; The memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the transfer model training method of any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the transfer model training method according to any one of claims 1 to 5.

Images