AI calculation training system

Abstract

The invention discloses an AI calculation training system which comprises an AI training platform, and the AI training platform comprises a data set management module which is used for creating a data set, labeling picture data or video data and screening data in the data set; the model training module is used for a user to select a model application type, a training data set, a training label and an adjusting and optimizing mode in a model training stage, and the system realizes model automatic training through an automatic deep learning technology; the model management module is used for managing various models and screening, deleting or checking data models; and the model verification module is used for helping a user to simulate an actual production environment through a model verification function and further evaluating a model effect. According to the method, a full-process graphical operation interface from data management to model release is provided, a user can complete model development work without any coding capability, the AI application development period is remarkably shortened, and the requirement for quick online application is met.

Description

Technical Field

[0001] The present invention relates to the field of AI technology, and in particular to an AI computing training system. Background Art

[0002] The application of AI technology in the industry field reflects the characteristics of in-depth scenarios and customized services. For industry users, on the one hand, they hope to use AI technology to improve efficiency and innovate business models; on the other hand, AI development faces a high technical threshold. For most industry users, the input-output ratio of establishing an AI technical team and developing related systems and applications is not high, and the engineering capabilities required for the implementation of scenarios are also a huge challenge. Based on research and summary, user pain points mainly include the following aspects: 1) There is a contradiction between demand and supply. Through the above analysis of typical intelligent needs of users, it can be found that the demand is scenario-based and fragmented. The business scenarios of various industries are different, and there are a large number of personalized needs, which require high development resource investment. For manufacturers, the needs of a single scenario lack universality, high single-point costs, and long payback cycles. Demand and supply cannot match, resulting in a large number of industries. The intelligent needs are difficult to meet and the intelligent upgrade is slow; 2) Traditional industries are eager to promote management mechanism changes and business efficiency improvements through AI technology. However, the high technical threshold and application cost of AI technology have become huge challenges. These challenges include high hardware investment, lack of high-quality training data, insufficient personnel skills reserves, and complex system integration. 3) Lack of high-quality training data. Image data is the basis for the development of AI intelligent perception algorithms. The quality and scale of training data play a key role in algorithm performance. The acquisition of image data required for current algorithm training generally encounters two problems: the currently deployed cameras only consider traditional security needs, and the collected images cannot fully match the pixels, angles and other requirements required for intelligent analysis tasks; limited by factors such as information security, trade secrets, and user privacy, some industry users find it difficult to provide large-scale effective data; 4) The cost of algorithm development is high. AI algorithm development has a high technical threshold. The AI ​​talent training cycle is long. In the short term, it is difficult for traditional companies to form a professional AI algorithm development team, and introducing it from outside also requires high labor costs. In addition, a large amount of data preprocessing and training tasks in the AI ​​development process require high-performance computing power and storage resources. Building a private heterogeneous computing cluster will greatly increase the IT burden of enterprises; 5) Intelligent hardware development is complex. The application of AI requires intelligent hardware that adapts to the scene. From the perspective of computing mode, it includes cloud reasoning and edge reasoning. Enterprises need to choose a reasonable deployment plan based on factors such as real-time performance, data security, and system scalability. After determining the plan, they also need to choose a suitable reasoning platform for AI algorithm model transplantation and optimization to give full play to computing power performance. This is also a task with high technical barriers; 6) The application integration delivery cycle is long. The essence of AI application is enterprise digital application. AI capabilities need to be seamlessly integrated with enterprise business systems to truly play a role in improving management models and improving business efficiency. On the other hand, AI intelligent hardware also needs to achieve maximum compatibility with the company's existing IT investment to achieve significant application benefits and return on investment at a lower cost.To this end, we propose an AI computing training system. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide an AI computing training system to solve the existing problems.

[0004] To achieve the above object, the present invention provides the following technical solutions: An AI computing training system includes an AI training platform, wherein the AI ​​training platform includes: The data set management module is used to create data sets, annotate image data or video data, and then filter the data in the data set; The model training module is used in the model training phase. Users select the model application type, training data set, training label and tuning mode. The system uses automatic deep learning technology to achieve automatic model training. The model management module is used to manage multiple models and filter, delete or view data models; The model verification module is used to help users simulate the actual production environment through the model verification function to further evaluate the model effect; The model publishing module is used to publish models approved by users.

[0005] Preferably, the data management module includes a data management module, a data annotation module, an intelligent annotation module, and a data quality diagnosis module. Preferably, the data management module includes: Dataset creation: Dataset types include training sets and verification sets. The data in the training set is used for model training; the data in the verification set is used for model effect verification and cannot be changed after use. Data upload is used to upload data in batches through compressed packages and folders. Preferably, the data annotation module includes: Image annotation: Image annotation methods include rectangle annotation, quadrilateral annotation, polygon annotation, line segment annotation and circle annotation; Video annotation is applicable to data annotation tasks of two model types, namely video behavior analysis and video classification. For the video behavior analysis model, it supports rectangular box annotation and labeling of consecutive frames of multiple targets in the video image, and for the video classification model, it supports labeling of video clips. The intelligent labeling module automatically completes data labeling through system-assisted manual labeling, reducing the workload of manual labeling and improving labeling efficiency. It can be divided into two modes: active learning and pre-labeling. Active learning means that the system actively learns and trains the user's unlabeled data based on the labels in the current labeling stage, and completes the labeling task through multiple rounds of iterations; Pre-labeling means that users can select a model that has been trained for the same task, quickly label all the data to be labeled, and complete the labeling task through manual confirmation. Preferably, the data quality diagnosis module includes an image quality diagnosis module and a target quality diagnosis module. Through data quality diagnosis, the user can adjust and optimize the image acquisition scheme and the annotation scheme according to the diagnosis results to assist in improving the model training effect; The image quality diagnosis module is used to analyze the original image data uploaded by users. By counting the distribution of image resolution and aspect ratio, it can diagnose image problems that may lead to poor model training results. The target quality diagnosis module is used to analyze the target boxes in the labeled data. By counting the target box area ratio, aspect ratio, density, and label distribution, it diagnoses labeling problems that may lead to poor model training results.

[0006] Preferably, the model training module includes: Datasets and labels are used to select labeled datasets as training samples. Users can flexibly select the labeled labels that need to be trained in the dataset, and the amount of labeled data required to start training varies according to different model types. Application type refers to the target deployment platform selected by the user. The system automatically selects the appropriate pre-trained model based on the computing power, memory, and power consumption of the deployment platform. Application types include cloud services, AI cameras, AI super brains, and AI servers. High-precision mode: For training tasks whose application type is cloud service, high-precision mode is supported. The system will automatically select a more complex pre-trained model to obtain better model performance. Intelligent model tuning supports transfer learning technology through the system, supports rapid modeling of small sample data for user scenarios, generates optimal models, and achieves rapid improvement of model accuracy and generalization performance.

[0007] Preferably, the model management module includes: The model version management module provides visual management of model projects and multiple iterations or deployable model versions under model projects; The system automatically randomly selects 10% of the training data for model effect verification, and automatically generates a performance evaluation report based on the verification results when the model training task is completed; Model Authorization Module In order to protect the user's algorithm model results, the system supports the model authorization mechanism.

[0008] Preferably, the model verification module includes: Local upload data: Upload original images in batches locally and perform batch verification. The system supports outputting verification results for each image. Create a validation set: Create a validation set in the data set management module and complete the validation set data annotation by manual annotation or intelligent annotation. The system supports the verification of the model through the validation set. After the model verification is completed, the model performance evaluation report is automatically generated.

[0009] Preferably, the model publishing module includes: Cloud Service API: For trained models whose application type is "cloud service", it is supported to publish the cloud service API; Model Download: For models whose application types have been trained except for "cloud service", it supports exporting model files online to local devices, and further deploying the models to AI cameras, AI super brains, and data center intelligent servers through the model deployment platform.

[0010] Preferably, the AI ​​training platform includes: Image perception models include object detection models, image single-label classification models, image multi-label classification models, image multi-attribute classification models, image instance segmentation models, image semantic segmentation models, hybrid models, image text recognition models, image comparison models, and retrieval comparison models; video perception models include video classification models and video behavior analysis models; Multimedia perception models include speech classification models.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The AI ​​training platform of the present invention is positioned for industry users and partners with zero algorithm foundation, aiming to lower the threshold and cost for users to use AI and corresponding services. This invention adopts AutoDL automatic learning technology and provides a full-process graphical operation interface from data management to model publishing. Users can complete model development without any coding ability, which significantly shortens the AI ​​application development cycle and meets the needs of rapid application launch. The present invention can enrich visual scenes and achieve high-precision model effects; it presets a wealth of basic algorithms and pre-trained models, and supports 10 visual perception scenarios; through incremental learning and transfer learning techniques, it helps users to train and obtain high-performance scenario-based models based on small sample data. Open cooperation is the new normal in the digital age. Hikvision builds the AIHub model and application recommendation platform based on the AI ​​open platform, allowing more excellent algorithms and solutions to be displayed, allowing more service providers to participate in the construction of industry digital transformation, allowing more industry users to enjoy the era dividends brought by digitalization, and jointly build an industry intelligent application ecosystem. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0013] Figure 1 It is a schematic diagram of the structure of the present invention; Figure 2 This is a schematic diagram of the functional architecture of the present invention; Figure 3 This is a standard schematic diagram of the picture of the present invention; Figure 4 This is a schematic diagram of intelligent annotation of the present invention; Figure 5 This is a schematic diagram of active learning of the present invention; Figure 6 This is a schematic diagram of image quality diagnosis of the present invention; Figure 7 This is a schematic diagram of target quality diagnosis of the present invention; Figure 8 This is a schematic diagram of the model training of the present invention; Fig. 9 This is a schematic diagram of model version management of the present invention; Fig.10 It is a schematic diagram of the performance evaluation report of the present invention; Fig.11 Schematic diagram of authorization for the model of the present invention: Fig.12 Creating a checksum schematic for the present invention; Fig.13 This is a schematic diagram of the cloud service API of the present invention; Fig.14 It is a schematic diagram of the model type of the present invention; Fig.15 It is a schematic diagram of object detection of the present invention; Fig.16 This is a schematic diagram of single-label classification of images according to the present invention; Fig.17 This is a schematic diagram of multi-label classification of images according to the present invention; Fig.18 This is a schematic diagram of image multi-attribute classification according to the present invention; Fig.19 It is a schematic diagram of image instance segmentation of the present invention; Fig. 20 This is a schematic diagram of image semantic segmentation of the present invention; Fig.21It is a schematic diagram of the hybrid model of the present invention; Fig. 22 This is a schematic diagram of image text recognition according to the present invention; Fig.23 It is a schematic diagram of image comparison of the present invention; Fig.24 A schematic diagram of the search and comparison of the present invention; Fig.25 A schematic diagram of the search and comparison of the present invention; Fig.26 This is a schematic diagram of video classification of the present invention; Fig. 27 This is a schematic diagram of speech classification of the present invention. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0015] like Figure 1-27 As shown, an AI computing training system includes an AI training platform, and the AI ​​training platform includes: a data set management module is used to create a data set, annotate image data or video data, and then filter the data in the data set.

[0016] like Figure 1 and Figure 2 As shown, the data management module includes a data management module, a data labeling module, an intelligent labeling module, and a data quality diagnosis module. The data management module comprises: Dataset creation: Dataset types include training sets and verification sets. The data in the training set is used for model training; the data in the verification set is used to verify the model effect and cannot be changed after use. Data upload: Support batch data upload through compressed packages, folders, etc. The image file format supports bmp, png, jpg, and jpeg; the video file format supports mp4; Supports online import and parsing of annotation files. It is suitable for scenarios where users complete data annotation and generate annotation files through offline annotation tools. Currently, it only supports the import of annotation files generated by offline annotation tools integrated in Hikvision C / S clients. Data Filtering: Supports filtering data in a dataset based on labeled tag names, allowing you to quickly locate and view data.

[0017] like Figure 3 As shown, the data annotation module includes: Image annotation: Image annotation methods include rectangle annotation, quadrilateral annotation, polygon annotation, line segment annotation and circle annotation. Rectangle is the most common and cheapest annotation method; quadrilateral annotation is suitable for the annotation task of the detection object with a large tilt angle in the image, which can effectively fit the detection target and avoid interference from useless background information; polygon, line segment and circle are only suitable for image semantic segmentation model type. Table 1 Marking method Video annotation is applicable to data annotation tasks of two model types, namely video behavior analysis and video classification. For the video behavior analysis model, it supports rectangular box annotation and labeling of consecutive frames of multiple targets in the video image, and for the video classification model, it supports labeling of video clips. The intelligent labeling module automatically completes data labeling through system-assisted manual labeling, reducing the workload of manual labeling and improving labeling efficiency. It can be divided into two modes: active learning and pre-labeling. Active learning means that the system actively learns and trains the user's unlabeled data based on the labels in the current labeling stage, and completes the labeling task through multiple rounds of iterations. The process is as follows: (1) Start intelligent labeling: manually label a small amount of data, with no less than 40 labeled images for each label; (2) The system screens and annotates some data: The system screens out images that need to be annotated first in batches. The system will start the next round of screening only after the currently screened images are manually confirmed (based on the system's pre-annotation results); (3) End intelligent labeling: Automatic end: The system determines that the current labeled images are sufficient and prompts you to end (the model trained with the current labeled images has almost the same effect as the model trained after all the data are labeled); Manual end: When the user confirms the filtered pictures, if he finds that the system pre-annotated boxes are very accurate, he can actively end the intelligent labeling. Pre-labeling means that users can select a model that has been trained for the same task, quickly label the current data to be labeled, and complete the labeling task through manual confirmation. The feature is that the effect of full labeling can be achieved with a small amount of labeling, and active learning is suitable for object detection models.

[0018] like Figure 6 and Figure 7As shown, the data quality diagnosis module includes an image quality diagnosis module and a target quality diagnosis module. Through data quality diagnosis, users can adjust and optimize image acquisition and annotation schemes according to the diagnosis results to help improve the model training effect. The image quality diagnosis module is used to analyze the original image data uploaded by the user, and diagnose image problems that may cause poor model training effects by statistically analyzing the distribution of image resolution and aspect ratio. Table 2 Image quality diagnosis The target quality diagnosis module is used to analyze the target boxes in the labeled data. By counting the target box area ratio, aspect ratio, density, and label distribution, it diagnoses labeling problems that may lead to poor model training results. Table 3 Target quality diagnosis

[0019] like Figure 8 As shown, the model training module is used in the model training phase. The user selects the model application type, training data set, training label and tuning mode. The system realizes automatic model training through automatic deep learning technology. The model training module includes: Datasets and labels are used to select labeled datasets as training samples. Users can flexibly select the labeled labels that need to be trained in the dataset, and the amount of labeled data required to start training varies depending on the model type. Application type refers to the target deployment platform selected by the user. The system automatically selects the appropriate pre-trained model based on the computing power, memory, and power consumption corresponding to the deployment platform. Application types include cloud services, AI cameras, AI super brains, and AI servers. High-precision mode For training tasks whose application type is cloud service, high-precision mode is supported. The system will automatically select a more complex pre-trained model to obtain better model performance. Intelligent model tuning supports transfer learning technology through the system, supports rapid modeling of small sample data for user scenarios, generates optimal models, and achieves rapid improvement of model accuracy and generalization performance.

[0019] like Fig. 9 As shown, the model management module is used to manage multiple models and to filter, delete or view data models. The model management module includes: The model version management module provides visual management of model projects and multiple iterative or deployable model versions under model projects. Supports viewing the model project list and filtering and viewing according to the model application type (target deployment platform); Supports model deletion. When deleting a model, all training versions under the model will be deleted synchronously. You can view model details, including version number, training start time, associated datasets, model status (including published, training, training completed, training stopped), model performance (trained models), and cloud service API URL (published models).

[0020] like Fig.10 As shown, the system automatically randomly selects 10% of the training data for model effect verification, and automatically generates a performance evaluation report based on the verification results when the model training task is completed. For training tasks of different model types, the system provides corresponding evaluation indicators and optimization suggestions to help users gain a preliminary understanding of model performance and decide whether to continue model iteration. The following lists the indicators and descriptions of the three model types: object detection, image classification, and image text recognition: Table 4 Index items

[0021] like Fig.11 As shown in the figure, the model authorization module protects the user's algorithm model results. The system supports the model authorization mechanism. After selecting model authorization, the model is automatically encrypted during the training phase. The system provides an authorization file generation service. During the model deployment phase, it supports the generation of license authorization files based on the user-specified device sequence, and imports them into the device together with the model file to ensure the normal operation of the model. The model verification module is used to help users simulate the actual production environment through the model verification function and further evaluate the model effect.

[0022] like Fig.12 As shown, the model verification module includes: Local upload data: Upload original images in batches locally and perform batch verification. The system supports outputting verification results for each image. Create a validation set: Create a validation set in the data set management module and complete the validation set data annotation by manual annotation or intelligent annotation. The system supports the verification of the model through the validation set. After the model verification is completed, the model performance evaluation report is automatically generated.

[0023] like Fig.13 As shown, the model publishing module is used to publish the model approved by the user. The model publishing module includes: Cloud Service API: For trained models whose application type is "cloud service", it is supported to publish the cloud service API; Model Download: For models whose application types have been trained except for "cloud service", it supports exporting model files online to local devices, and further deploying the models to AI cameras, AI super brains, and data center intelligent servers through the model deployment platform.

[0024] like Fig.14 As shown, the AI ​​training platform includes: Image perception models include object detection models, image single-label classification models, image multi-label classification models, image multi-attribute classification models, image instance segmentation models, image semantic segmentation models, hybrid models, image text recognition models, image comparison models, and retrieval comparison models; The video perception model includes the video classification model and the video behavior analysis model; Multimedia perception models include speech classification models.

[0025] like Fig.15 As shown, the object detection model is tailored to identify the location and category of each object in the image.

[0026] like Fig.16 As shown in the figure, the image single-label classification model can be customized to identify whether the image contains certain characteristic objects, states or scenes. Among them, the image single classification is suitable for analyzing the requirements of only one attribute of the image. For example, only cloud, rime and frost are classified.

[0027] like Fig.17 As shown in the figure, the image multi-label classification model gives an image and correctly gives the category to which the image belongs. It is suitable for analyzing the needs of images with multiple attributes at the same time.

[0028] like Fig.18 As shown in the figure, the image multi-attribute classification model is customized to identify multiple objects and states in the image. For example, scene content analysis, whether the image contains sky, grass, buildings, trees, etc. The multi-label classification function is suitable for tasks where the analysis object has multiple objects and states, and each analysis sample can output multiple labels.

[0029] like Fig.19 As shown in the figure, compared with semantic segmentation, the instance segmentation model can output the category and position (contour) of each target in the image, as well as the number of targets in each category, while semantic segmentation can only output the contour information of each category in the image and cannot distinguish different targets.

[0030] like Fig. 20As shown, the image semantic segmentation model classifies images at the pixel level and is suitable for identifying image regions of each category in the picture. Typical application scenarios include: irregular object detection (such as river floating object identification, oil leak detection, foam detection), defect detection (such as oil leak detection, workpiece damage), fine contour (human body part segmentation, facial features segmentation, meter pointer segmentation), target area calculation, etc.

[0031] like Fig.21 As shown, the mixed model: Compared with the object detection model, the hybrid model can identify the attributes of the target and perform detailed identification of the target's attribute information. The hybrid model can effectively solve the following three types of problems: (1) Classify specific moving areas: First, locate the specific area through the detection model, and then perform secondary classification on the area. For example: Identify the status of the iron plate of the production line trolley. (2) The number of target types that the object detection model cannot detect is greater than 16: The detection model can be used to locate the target object, and then classification can be used to extract the target category attributes. For example: office supplies recognition. (3) Using object detection models results in a high misidentification rate: The detection model can be used to locate the target, detect it as one category, and then classify its attributes. For example: detecting whether a person is wearing a helmet.

[0032] like Fig. 22 As shown in the figure, the image text recognition model is suitable for custom recognition of the position and content of each string in the picture of natural scenes. It supports English characters, numbers, Chinese characters and common character combinations. For example: carriage number recognition, food packaging number recognition, etc. Note: It is not suitable for card ticket and printed document recognition that requires information structuring.

[0033] like Fig.23 As shown in the figure, the image comparison model is suitable for identifying the difference area between the current image and the reference image in the same scene. Compared with object detection, the advantages are: 1) it can solve the open set problem that object detection cannot achieve (that is, the objects to be analyzed cannot be exhaustive); 2) it can solve the problem of irregular object detection. Typical scenarios include: equipment and facility status recognition (door opening and closing recognition), oil leakage (irregular object) recognition, channel occupancy detection, etc.

[0034] like Fig.24 As shown, the retrieval and comparison model is used to customize the target subject in the recognition image, and further identify the category of the target subject by retrieving and comparing the similarity between the registration set and the target subject.

[0035] like Fig.25As shown in the figure, retrieval comparison realizes the detection and recognition capability of a large number of target categories in a certain business scenario. Compared with object detection, the advantages are: 1) it can recognize more target categories (usually the number of categories is more than 1,000); 2) the model has a certain generalization ability. When a new target category is added, there is no need to retrain the model using the new target category as a training sample. It is suitable for business scenarios where target categories are frequently added or updated. Typical scenarios include: warehouse material identification, shelf display status analysis, etc.

[0036] like Fig.26 As shown in Figure 1, the video classification model is suitable for identifying object attributes or scene attributes in a video sequence. Compared with the video behavior analysis model, it does not contain time point and spatial position information.

[0037] like Fig. 27 As shown in the figure, the video behavior analysis model is suitable for customizing the category, time period and spatial position of a behavior in a video clip. It can analyze video clips to solve the needs of time-series analysis that cannot be confirmed from a single picture. It is divided into two steps: detection and positioning and behavior classification. Target positioning requires the position of the object to be analyzed in the time sequence through detection and tracking, and behavior classification is performed on a sequence to output the analysis results. Typical scenarios include: 1) Individual behavior analysis: such as falling to the ground, raising hands, bending over, scratching the head, etc.; 2) Individual behaviors of human-object interaction: such as playing with mobile phones, dragging boxes, etc. Human-to-human interactions such as hugging and fighting are not supported yet.

[0038] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI computing training system, characterized in that: An AI training platform is included, and the AI ​​training platform includes: The data set management module is used to create data sets, annotate image data or video data, and then filter the data in the data set; The model training module is used in the model training phase. Users select the model application type, training data set, training label and tuning mode. The system uses automatic deep learning technology to achieve automatic model training. The model management module is used to manage multiple models and filter, delete or view data models; The model verification module is used to help users simulate the actual production environment through the model verification function to further evaluate the model effect; The model publishing module is used to publish models approved by users.

2. The AI ​​computing training system according to claim 1, characterized in that: The data management module includes a data management module, a data annotation module, an intelligent annotation module, and a data quality diagnosis module.

3. The AI ​​computing training system according to claim 1, characterized in that: The data management module comprises: Dataset creation: Dataset types include training sets and verification sets. The data in the training set is used for model training; the data in the verification set is used for model effect verification and cannot be changed after use. Data upload is used to upload data in batches through compressed packages and folders.

4. The AI ​​computing training system according to claim 1, characterized in that: The data annotation module includes: Image annotation: Image annotation methods include rectangle annotation, quadrilateral annotation, polygon annotation, line segment annotation and circle annotation; Video annotation is applicable to data annotation tasks of two model types, namely video behavior analysis and video classification. For the video behavior analysis model, it supports rectangular box annotation and labeling of consecutive frames of multiple targets in the video image, and for the video classification model, it supports labeling of video clips. The intelligent labeling module automatically completes data labeling through system-assisted manual labeling, reducing the workload of manual labeling and improving labeling efficiency. It can be divided into two modes: active learning and pre-labeling. Active learning means that the system actively learns and trains the user's unlabeled data based on the labels in the current labeling stage, and completes the labeling task through multiple rounds of iterations; Pre-labeling means that users can select a model that has been trained for the same task, quickly label all the data to be labeled, and complete the labeling task through manual confirmation.

5. The AI ​​computing training system according to claim 1, characterized in that: The data quality diagnosis module includes an image quality diagnosis module and a target quality diagnosis module. Through data quality diagnosis, users can adjust and optimize image acquisition and annotation schemes according to the diagnosis results to help improve the model training effect; The image quality diagnosis module is used to analyze the original image data uploaded by users. By counting the distribution of image resolution and aspect ratio, it can diagnose image problems that may lead to poor model training results. The target quality diagnosis module is used to analyze the target boxes in the labeled data. By counting the target box area ratio, aspect ratio, density, and label distribution, it diagnoses labeling problems that may lead to poor model training results.

6. The AI ​​computing training system according to claim 1, characterized in that: The model training module includes: Datasets and labels are used to select labeled datasets as training samples. Users can flexibly select the labeled labels that need to be trained in the dataset, and the amount of labeled data required to start training varies according to different model types. Application type refers to the target deployment platform selected by the user. The system automatically selects the appropriate pre-trained model based on the computing power, memory, and power consumption of the deployment platform. Application types include cloud services, AI cameras, AI super brains, and AI servers. High-precision mode: For training tasks whose application type is cloud service, high-precision mode is supported. The system will automatically select a more complex pre-trained model to obtain better model performance. Intelligent model tuning supports transfer learning technology through the system, supports rapid modeling of small sample data for user scenarios, generates optimal models, and achieves rapid improvement of model accuracy and generalization performance.

7. The AI ​​computing training system according to claim 1, characterized in that: The model management module includes: The model version management module provides visual management of model projects and multiple iterations or deployable model versions under model projects; The system automatically randomly selects 10% of the training data for model effect verification, and automatically generates a performance evaluation report based on the verification results when the model training task is completed; Model Authorization Module In order to protect the user's algorithm model results, the system supports the model authorization mechanism.

8. The AI ​​computing training system according to claim 1, characterized in that: The model verification module includes: Local upload data: Upload original images in batches locally and perform batch verification. The system supports outputting verification results for each image. Create a validation set: Create a validation set in the data set management module and complete the validation set data annotation by manual annotation or intelligent annotation. The system supports the verification of the model through the validation set. After the model verification is completed, the model performance evaluation report is automatically generated.

9. The AI ​​computing training system according to claim 1, characterized in that: The model publishing module includes: Cloud Service API: For trained models whose application type is "cloud service", it is supported to publish the cloud service API; Model Download: For models whose application types have been trained except "cloud service", it supports exporting model files online to local devices, and further deploying the models to AI cameras, AI super brains, and data center intelligent servers through the model deployment platform.

10. The AI ​​computing training system according to claim 1, characterized in that: The AI ​​training platform includes: Image perception models include object detection models, image single-label classification models, image multi-label classification models, image multi-attribute classification models, image instance segmentation models, image semantic segmentation models, hybrid models, image text recognition models, image comparison models, and retrieval comparison models; The video perception model includes the video classification model and the video behavior analysis model; Multimedia perception models include speech classification models.

Images