A training system for artificial intelligence models
Through the training system of artificial intelligence model, including training set generation module, parameter configuration module and model evaluation module, the problem of high difficulty in developing artificial intelligence model training and debugging is solved by scientific researchers in the medical field, and the rapid generation is achieved according to user needs and adjust the model to improve the ease of training and debugging.
Patent Information
- Application Number
- CN202210333808.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-03-31
AI Technical Summary
In the prior art, scientific researchers in the medical field need to make significant adjustments to the algorithms and codes during the training and debugging of artificial intelligence models, which leads to high development difficulties and relying on algorithm engineers, which makes the process inconvenient.
Provides a training system for artificial intelligence models, including training set generation module, parameter configuration module and training module, generates specific types of training sets and training parameters according to the user's model needs, and evaluates the trained model through the model evaluation module. and adjust, generate correction parameters to meet user needs.
By setting up the training set generation module and the training parameter configuration module, users can directly generate the corresponding training set and training parameters according to their own needs, reducing the difficulty of developing artificial intelligence models and improving the simplicity of training and debugging.
Smart Images

Figure CN114694009B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence training technology, and in particular to a training system for an artificial intelligence model. Background Art
[0002] In the prior art, there are technical solutions that use different types of artificial intelligence models in medical ultrasound examinations to assist in the identification of ultrasound images and generate corresponding annotation information. This type of technical solution usually sets corresponding training sets for different types of image features that need to be identified, and adjusts the parameters of the artificial intelligence model to obtain the artificial intelligence model required by the user. However, in the actual implementation process, the inventors found that in the above-mentioned debugging and training process, it is necessary to make significant adjustments to the relevant algorithms and codes, and many scientific researchers in the medical field have weak development capabilities and it is difficult to adjust the algorithms and codes of artificial intelligence models, which causes certain inconveniences to the training and debugging of artificial intelligence models. Summary of the invention
[0003] In view of the above problems existing in the prior art, a training system for an artificial intelligence model is now provided.
[0004] The specific technical solutions are as follows:
[0005] A training system for an artificial intelligence model, comprising:
[0006] The training set generation module processes the training data according to an external input model requirement to generate a training set;
[0007] A parameter configuration module, wherein the parameter configuration module generates training configuration parameters according to an externally input training control instruction;
[0008] A training module, wherein the training module is connected to the training set generation module and the parameter configuration module, wherein the training module trains the artificial intelligence model according to the training set and the training configuration parameters, and then outputs the trained artificial intelligence model.
[0009] Preferably, it also includes:
[0010] A model evaluation module, the model evaluation module is connected to the training module and obtains the trained artificial intelligence model;
[0011] The model evaluation module evaluates the artificial intelligence model according to the model requirements to generate an evaluation result, and then feeds back correction parameters to the training module according to the evaluation result.
[0012] Preferably, the training set generation module includes:
[0013] A data storage submodule, wherein the training data is pre-stored in the data storage submodule;
[0014] A training set encapsulation submodule, wherein the training set encapsulation submodule is connected to the data storage submodule and packages the training data into the training set;
[0015] A contour scaling submodule, the contour scaling submodule is connected to the training set encapsulation submodule, and the contour scaling submodule scales the contour size of the training data in the training set according to the model requirements;
[0016] A contour conversion submodule, the contour conversion submodule is connected to the training set encapsulation submodule, and the contour conversion submodule adjusts the contour of the training data in the training set to a convex hull contour according to the model requirements;
[0017] A sample selection submodule, wherein the sample selection submodule is connected to the training set encapsulation submodule, and the sample selection submodule selects a specific type of the training data from the training data according to the model requirements, and adds the specific type of the training data to the training set to increase the proportion of the specific type of the training data in the training set.
[0018] Preferably, the parameter configuration module includes:
[0019] an optimizer setting submodule, wherein the optimizer setting submodule sets a corresponding optimizer in the training configuration parameters according to the training control instruction;
[0020] A training parameter setting submodule, wherein the training parameter setting submodule sets the training parameters in the training configuration parameters according to the training control instruction;
[0021] The training parameters include learning rate, number of training times and number of batches.
[0022] Preferably, the training module includes:
[0023] A model training submodule, wherein the model training submodule is connected to the training set generation module and the parameter configuration module, and the model training submodule trains the artificial intelligence model according to the training configuration parameters and the training set;
[0024] A process display submodule, wherein the process display submodule is connected to the model training submodule, and the process display submodule obtains the operating data generated by the model training submodule during the training process, and displays the operating data through a visual interface.
[0025] Preferably, the model evaluation module comprises:
[0026] A model testing submodule, wherein the model testing submodule is connected to the training module, and the model testing submodule tests the trained artificial intelligence model according to the model requirements and outputs the evaluation result;
[0027] An adjustment submodule, the adjustment submodule is connected to the model testing submodule, and the adjustment submodule determines whether the artificial intelligence model needs to be adjusted according to the evaluation result;
[0028] When the artificial intelligence model needs to be adjusted, the adjustment submodule generates the correction parameters according to the model requirements.
[0029] Preferably, the model requirements include: model specificity requirements, model sensitivity requirements, accuracy for specific samples, and balanced recognition requirements for all samples.
[0030] Preferably, when the model requirement is the model sensitivity requirement, the adjustment submodule extracts an erroneous negative sample from the evaluation result, and adjusts the sensitivity level according to the erroneous negative sample to generate the correction parameter;
[0031] When the model requirement is the model specificity requirement, the adjustment submodule extracts false positive samples from the evaluation results, and adjusts the sensitivity level according to the false positive samples to generate the correction parameters.
[0032] Preferably, the model evaluation module comprises:
[0033] An interaction submodule, the interaction submodule is connected to the model testing submodule, and the interaction submodule obtains and displays the evaluation result;
[0034] The interaction submodule receives an externally input adjustment instruction and controls the adjustment submodule to generate the correction parameter according to the adjustment instruction.
[0035] The above technical solution has the following advantages or beneficial effects: by setting up a training set generation module and a training parameter configuration module, users can directly generate corresponding training sets and training parameters according to their expected requirements for the model, and then control the direction of model training according to model requirements during the training process, thereby obtaining the desired artificial intelligence model, which improves the convenience of artificial intelligence model training and debugging. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The embodiments of the present invention will be described more fully with reference to the attached drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.
[0037] Figure 1It is an overall schematic diagram of an embodiment of the present invention;
[0038] Figure 2 Schematic diagram of a parameter configuration module in an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0041] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.
[0042] The present invention comprises:
[0043] A training system for an artificial intelligence model, such as Figure 1 As shown, including:
[0044] The training set generation module 1 processes the training data according to an external input model requirement to generate a training set;
[0045] Parameter configuration module 2, parameter configuration module 2 generates training configuration parameters according to an externally input training control instruction;
[0046] Training module 3, training module 3 connects training set generation module 1 and parameter configuration module 2, training module 3 trains the artificial intelligence model according to the training set and training configuration parameters, and then outputs the trained artificial intelligence model.
[0047] Specifically, in order to address the problem in the prior art that, during the artificial intelligence development process, it is necessary to rely on algorithm engineers to debug the artificial intelligence model, which makes the development process inconvenient, in this embodiment, a training set generation module 1 and a parameter configuration module 2 are set to generate a specific type of training set according to model requirements, so that the required model can be quickly obtained according to the model requirements during the training process, thereby reducing the difficulty of developing the artificial intelligence model.
[0048] During the implementation process, the training system is set in a computer device as a computer program. During the implementation process, users often have the following requirements for different types of recognition needs: model specificity, model sensitivity, special data samples or balanced recognition, etc. In response to the above requirements, the training set generation module 1 can adopt different training set generation strategies according to the corresponding requirements, so that the trained artificial intelligence model meets the user's expected requirements.
[0049] In a preferred embodiment, it also includes:
[0050] Model evaluation module 4, model evaluation module 4 is connected to training module 3 and obtains the trained artificial intelligence model;
[0051] The model evaluation module 4 evaluates the artificial intelligence model according to the model requirements to generate an evaluation result, and then feeds back the correction parameters to the training module 3 according to the evaluation result.
[0052] Specifically, in view of the problem that in the prior art, the development process of artificial intelligence needs to rely on algorithm engineers to adjust parameters of the trained artificial intelligence model, change algorithms, etc., which makes the development process inconvenient, in this embodiment, by setting a model evaluation module 4, the trained artificial intelligence model is verified, and the evaluation result is automatically generated, and then according to the model requirements, that is, the above-mentioned model specificity, model sensitivity, special data samples or balanced recognition, it is judged whether the trained artificial intelligence model is the artificial intelligence model required by the user. When the trained artificial intelligence model does not meet the user's requirements, the parameters of the trained artificial intelligence model are adjusted by generating correction parameters, so that the artificial intelligence model meets the user's expectations.
[0053] In a preferred embodiment, the training set generation module 1 includes:
[0054] A data storage submodule 11, in which training data is pre-stored;
[0055] The training set encapsulation submodule 12 is connected to the data storage submodule 11 and packages the training data into a training set;
[0056] The contour scaling submodule 13 is connected to the training set encapsulation submodule 12, and the contour scaling submodule 13 scales the contour size of the training data in the training set according to the model requirements;
[0057] The contour conversion submodule 14 is connected to the training set encapsulation submodule 12, and the contour conversion submodule 14 adjusts the contour of the training data in the training set to a convex hull contour according to the model requirements;
[0058] The sample selection submodule 15 is connected to the training set encapsulation submodule 12. The sample selection submodule 15 selects a specific type of training data from the training data according to the model requirements, and adds the specific type of training data to the training set to increase the proportion of the specific type of training data in the training set.
[0059] Specifically, in order to address the problem in the prior art that, during the artificial intelligence development process, it is necessary to rely on algorithm engineers to debug the artificial intelligence model, which makes the development process inconvenient, in this embodiment, by sequentially setting the contour scaling submodule 13, the contour conversion submodule 14, and the sample selection submodule 15 to be connected to the training set encapsulation submodule 12, the training data in the training set is adjusted according to the user's model requirements, so that the artificial intelligence model trained based on the training set meets the user's expectations for the model.
[0060] During the implementation process, the corresponding annotated training images are pre-stored in the data storage submodule 11, such as ultrasound images of specific parts, including normal ultrasound images of the parts and ultrasound images with lesions. The contours and types of the lesion sites have been pre-annotated. The user's model requirements include model specificity, model sensitivity, special data samples or balanced recognition. When the model requirement is balanced recognition, the training data is directly packaged and output as a training set only through the training set encapsulation submodule 12. When the model requirement is model specificity, the label contour is reduced by using the contour scaling submodule 13 to extract the core part of the lesion, thereby improving the specificity; when the model requirement is model sensitivity, the contour conversion submodule 14 is used to set the contour to a convex hull contour, so that the surrounding image is included in the recognition feature to improve the sensitivity of the model.
[0061] In a preferred embodiment, the parameter configuration module 2 includes:
[0062] The optimizer setting submodule 21 sets the corresponding optimizer in the training configuration parameters according to the training control instruction;
[0063] A training parameter setting submodule 22, the training parameter setting submodule 22 sets the training parameters in the training configuration parameters according to the training control instruction;
[0064] Training parameters include learning rate, number of training steps and number of batches.
[0065] Specifically, in view of the problem in the prior art that in the process of artificial intelligence development, it is necessary to rely on algorithm engineers to debug the artificial intelligence model, which makes the development process inconvenient, in this embodiment, by setting Figure 2The parameter configuration module 2 shown enables the user to set the optimizer, including Adam, SGD, RMSprop and other optimizers, and training parameters according to the training control instructions, thereby reducing the difficulty of development.
[0066] In a preferred embodiment, the training module 3 includes:
[0067] The model training submodule 31 is connected to the training set generation module 1 and the parameter configuration module 2, and the model training submodule 31 trains the artificial intelligence model according to the training configuration parameters and the training set;
[0068] The process display submodule 32 is connected to the model training submodule 31. The process display submodule 32 obtains the operation data generated by the model training submodule 31 during the training process and displays the operation data through a visual interface.
[0069] Specifically, in order to address the problem in the prior art that, during the artificial intelligence development process, it is necessary to rely on algorithm engineers to debug the artificial intelligence model, which makes the development process inconvenient, in this embodiment, the artificial intelligence model is trained by setting a model training submodule 31 in the training module 3, and the model training submodule 31 is connected through the process display submodule 32 to collect the operation data of the model training submodule 31 during the training process, so that the process display submodule 32 can provide real-time feedback on various performances of the model based on the operation data on the visual interface, thereby prompting the user with the training direction of the current model, so that the user can judge whether the training of the model meets expectations.
[0070] In a preferred embodiment, the model evaluation module 4 includes:
[0071] Model testing submodule 41, the model testing submodule 41 is connected to the training module, and the model testing submodule 41 tests the trained artificial intelligence model according to the model requirements and outputs the evaluation results;
[0072] An adjustment submodule 42, the adjustment submodule 42 is connected to the model testing submodule 41, and the adjustment submodule 42 determines whether the artificial intelligence model needs to be adjusted according to the evaluation result;
[0073] When the artificial intelligence model needs to be adjusted, the adjustment submodule generates correction parameters based on the model requirements.
[0074] Specifically, in view of the problem that in the prior art, the development process of artificial intelligence needs to rely on algorithm engineers to adjust parameters and change algorithms of the trained artificial intelligence model, which makes the development process inconvenient, in this embodiment, a model testing submodule 41 is set to verify the trained model using a verification set corresponding to the training set, so as to determine and generate an evaluation result. At the same time, an adjustment submodule 42 is set to automatically determine whether the trained model meets the user's model requirements based on the evaluation results, and generate correction parameters to adjust the model when the model does not meet the user's requirements, so that the model meets the user's requirements.
[0075] In a preferred embodiment, the model requirements include: model specificity requirements, model sensitivity requirements, accuracy for specific samples, and balanced recognition requirements for all samples.
[0076] In a preferred embodiment, when the model requirement is a model sensitivity requirement, the adjustment submodule 42 extracts an erroneous negative sample from the self-evaluation result, and adjusts the sensitivity level according to the erroneous negative sample to generate a correction parameter;
[0077] When the model requirement is a model specificity requirement, the adjustment submodule 42 extracts false positive samples from the evaluation results and adjusts the sensitivity level according to the false positive samples to generate correction parameters.
[0078] Specifically, in view of the problem that in the prior art, the development process of artificial intelligence needs to rely on algorithm engineers to adjust parameters and change algorithms of trained artificial intelligence models, which makes the development process inconvenient, in this embodiment, erroneous positive samples or erroneous negative samples are obtained according to different model requirements, so as to adjust the sensitivity level of the model. Specifically, when the model requirement is model sensitivity, the highlight visualization model infers the erroneous negative sample results that are higher than the threshold, and decides whether to lower or increase the current sensitivity level. When the sensitivity level needs to be increased, the sensitivity of the model is increased by expanding the contour outward by 5 pixels. In one embodiment, the sensitivity level can be increased by a maximum of 8 levels and 40 pixels. When the sensitivity level needs to be lowered, by lowering the sensitivity level and shrinking 5 pixels inward, it can be scaled to a maximum of 70% of the original.
[0079] In a preferred embodiment, the model evaluation module 4 further includes:
[0080] Interaction submodule 43, the interaction submodule 43 is connected to the model testing submodule 41, and the interaction submodule 43 obtains and displays the evaluation results;
[0081] The interaction submodule 43 receives an adjustment instruction input from the outside and controls the adjustment submodule 42 to generate a correction parameter according to the adjustment instruction.
[0082] The beneficial effect of the present invention is that by setting up a training set generation module and a training parameter configuration module, users can directly generate corresponding training sets and training parameters according to their own expected requirements for the model, and then control the direction of model training according to model requirements during the training process, thereby obtaining the desired artificial intelligence model, thereby improving the simplicity of artificial intelligence model training and debugging.
[0083] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.
Claims
1. A training system for an artificial intelligence model, characterized in that: include: A training set generation module, wherein the training set generation module processes the training data according to an external input model requirement to generate a training set; The model requirements are model specificity, model sensitivity, special data samples or equilibrium identification; A parameter configuration module, wherein the parameter configuration module generates training configuration parameters according to an externally input training control instruction; A training module, wherein the training module is connected to the training set generation module and the parameter configuration module, and the training module trains the artificial intelligence model according to the training set and the training configuration parameters, and then outputs the trained artificial intelligence model; The training data is an annotated ultrasound image of a specific part; The training set generation module comprises: A contour scaling submodule, the contour scaling submodule is connected to the training set encapsulation submodule, and the contour scaling submodule scales the contour size of the training data in the training set according to the model requirements; The contour scaling submodule scaling the contour size of the training data in the training set includes scaling the label contour of the training data to extract the core part of the lesion.
2. The training system according to claim 1, characterized in that Also includes: A model evaluation module, the model evaluation module is connected to the training module and obtains the trained artificial intelligence model; The model evaluation module evaluates the artificial intelligence model according to the model requirements to generate an evaluation result, and then feeds back correction parameters to the training module according to the evaluation result.
3. The training system according to claim 1, characterized in that The training set generation module comprises: A data storage submodule, wherein the training data is pre-stored in the data storage submodule; A training set encapsulation submodule, the training set encapsulation submodule is connected to the data storage submodule and the contour scaling submodule, and packages the training data into the training set; A contour conversion submodule, the contour conversion submodule is connected to the training set encapsulation submodule, and the contour conversion submodule adjusts the contour of the training data in the training set to a convex hull contour according to the model requirements; A sample selection submodule, wherein the sample selection submodule is connected to the training set encapsulation submodule, and the sample selection submodule selects a specific type of the training data from the training data according to the model requirements, and adds the specific type of the training data to the training set to increase the proportion of the specific type of the training data in the training set.
4. The training system according to claim 1, characterized in that The parameter configuration module includes: an optimizer setting submodule, wherein the optimizer setting submodule sets a corresponding optimizer in the training configuration parameters according to the training control instruction; A training parameter setting submodule, wherein the training parameter setting submodule sets training parameters in the training configuration parameters according to the training control instruction; The training parameters include learning rate, number of training times and number of batches.
5. The training system according to claim 1, characterized in that: The training module includes: A model training submodule, wherein the model training submodule is connected to the training set generation module and the parameter configuration module, and the model training submodule trains the artificial intelligence model according to the training configuration parameters and the training set; A process display submodule, wherein the process display submodule is connected to the model training submodule, and the process display submodule obtains the operating data generated by the model training submodule during the training process, and displays the operating data through a visual interface.
6. The training system according to claim 2, characterized in that: The model evaluation module includes: A model testing submodule, wherein the model testing submodule is connected to the training module, and the model testing submodule tests the trained artificial intelligence model according to the model requirements and outputs the evaluation result; An adjustment submodule, the adjustment submodule is connected to the model testing submodule, and the adjustment submodule determines whether the artificial intelligence model needs to be adjusted according to the evaluation result; When the artificial intelligence model needs to be adjusted, the adjustment submodule generates the correction parameters according to the model requirements.
7. The training system according to claim 6, characterized in that The model requirements include: model specificity requirements, model sensitivity requirements, accuracy for specific samples, and balanced recognition requirements for all samples.
8. The training system according to claim 7, characterized in that When the model requirement is the model sensitivity requirement, the adjustment submodule extracts an erroneous negative sample from the evaluation result, and adjusts the sensitivity level according to the erroneous negative sample to generate the correction parameter; When the model requirement is the model specificity requirement, the adjustment submodule extracts a false positive sample from the evaluation result, and adjusts the sensitivity level according to the false positive sample to generate the correction parameter.
9. The training system according to claim 6, characterized in that: The model evaluation module also includes: An interaction submodule, the interaction submodule is connected to the model testing submodule, and the interaction submodule obtains and displays the evaluation result; The interaction submodule receives an externally input adjustment instruction and controls the adjustment submodule to generate the correction parameter according to the adjustment instruction.
