Data development system

Abstract

The invention provides a data development system, and the system comprises an image obtaining module which obtains an original image data set in a format matched with a target visual task; the data analysis module splits the original image data set according to a preset splitting strategy to obtain a split image data set; the split image data set at least comprises a training image data set and a verification image data set; the model training module processes the training image data set according to a preset data enhancement mode, and performs model training based on an enhanced image data set obtained through processing to obtain a target model adaptive to the target visual task; the model evaluation module determines a verification result of the target model on the verification image data set. In the AI development realization process, different requirements of different visual tasks on model training are considered, model training with higher robustness is realized based on a data enhancement processing mode, and the whole development process does not need a high professional skill level, so that the development efficiency is remarkably improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a data development system. Background Art

[0002] Since the Dartmouth Conference in 1956, artificial intelligence (AI) technology has gone through the initial development period (1956-1960), the reflection development period (1960-1970), the application development period (1970-1980), the sluggish development period (1980-1990), the steady development period (1990-2010), and finally entered a period of vigorous development in 2011. With the rapid development of computer technology and the background of rapid development of global informatization, AI is leading a new round of scientific and technological innovation.

[0003] However, current AI technology has high requirements for developers. Data development based on artificial intelligence relies on the technical level of developers, with low development efficiency. In addition, most existing solutions are aimed at a specific scenario (such as mask recognition, liveness detection, pedestrian detection, etc.). After changing the project scenario, the code needs to be re-adapted and the parameters adjusted. Later project transplantation, maintenance, and iterative upgrades are difficult, and the overall compatibility is poor. Summary of the invention

[0004] The embodiments of the present disclosure at least provide a data development system to realize AI development for various visual tasks. The entire system has better compatibility and is simple to operate, which significantly improves development efficiency.

[0005] In a first aspect, an embodiment of the present disclosure provides a data development system, including:

[0006] An image acquisition module, configured to acquire an original image data set in a format matching the target visual task in response to a data acquisition instruction for the target visual task in the user interface;

[0007] A data analysis module, configured to respond to a splitting instruction for the original image data set and split the original image data set according to a preset splitting strategy to obtain a split image data set; the split image data set at least includes a training image data set and a verification image data set;

[0008] A model training module, for responding to a model training instruction for the target visual task, processing the training image data set according to a preset data enhancement method, and performing model training based on the enhanced image data set obtained by the processing to obtain a target model adapted to the target visual task;

[0009] The model evaluation module is used to determine the verification result of the target model on the verification image dataset in response to the evaluation instruction for the target model.

[0010] In a possible implementation, the model training module includes a data enhancement unit;

[0011] The data enhancement unit is used to randomly extract a data enhancement method from the enhancement space and randomly extract an enhancement amplitude under the data enhancement method in response to a trigger instruction for the automatic data enhancement strategy under the target visual task; use the randomly extracted data enhancement method and its enhancement amplitude to process part or all of the image data in the training image data set to obtain an enhanced image data set.

[0012] In a possible implementation, the model training module includes a first hyperparameter search unit;

[0013] The first hyperparameter search unit, in response to the first hyperparameter search instruction for the target visual task, selects a target hyperparameter type that matches the target visual task from various hyperparameter types based on the task type of the target visual task; obtains an initialized hyperparameter value corresponding to the target hyperparameter type, and cyclically searches and updates the hyperparameter value based on a preset Bayesian hyperparameter search method until a preset search cutoff condition is reached, thereby obtaining the searched hyperparameter value.

[0014] In a possible implementation, the first hyperparameter search unit is specifically used to select a target hyperparameter type that matches the original image dataset from various hyperparameter types based on the task type of the target visual task and the original image dataset acquired under the target visual task.

[0015] In a possible implementation, the model training module includes a second hyperparameter search unit;

[0016] The second super-parameter search unit is used to locally obtain the hyper-parameters customized by the user under the second super-parameter search instruction in response to the user's second super-parameter search instruction for the target visual task.

[0017] In a possible implementation, the model training module includes a third hyperparameter search unit;

[0018] The third super-parameter search unit is used to respond to the expert's third super-parameter search instruction for the target visual task, and remotely obtain the hyper-parameters set with the expert's assistance under the third super-parameter search instruction.

[0019] In a possible implementation, the model training module includes a model training unit;

[0020] The model training unit is used to obtain a target model to be trained for the target visual task, input the processed enhanced image data set into a pre-trained target model for at least one round of model training, and obtain a target model suitable for the target visual task.

[0021] In a possible implementation manner, the model training unit is specifically used to:

[0022] Obtain multiple candidate models corresponding to each visual task;

[0023] Determining a candidate model corresponding to the target visual task from the multiple candidate models based on the task type of the target visual task;

[0024] The determined candidate model is used as a target model to be trained for the target visual task.

[0025] In a possible implementation, it also includes an interface display module;

[0026] The interface display module is used to display the verification result in response to a display request of the target user for the verification result.

[0027] In a possible implementation manner, the interface display module is further used to:

[0028] In response to a target user's reasoning request for the target model, at least one image data input by the user is obtained, and the at least one image data is sequentially input into the target model for model reasoning, a reasoning evaluation result is determined, and the reasoning evaluation result is displayed.

[0029] In a possible implementation, it further includes:

[0030] The one-click upload module is used to upload the target model to the model cloud store bound to the target user account with one click when it is determined that the reasoning evaluation result meets the preset reasoning requirements, so that the required model can be obtained at any time through the model cloud store.

[0031] In a possible implementation manner, the image acquisition module is specifically used to:

[0032] Extracting image data matching the task type of the target visual task according to a preset multi-level data set construction structure;

[0033] The original image data set is constructed based on the extracted image data.

[0034] By adopting the above data development system, the user can realize AI development through the user interface through the mutual cooperation of the image acquisition module, the data analysis module, the model training module and the model evaluation module, wherein the image acquisition module obtains the original image data set in a format that matches the target visual task; the data analysis module splits the original image data set according to the preset splitting strategy to obtain the split image data set; the model training module processes the split training image data set according to the preset data enhancement method, and performs model training based on the enhanced image data set obtained by the processing to obtain the target model adapted to the target visual task; the model evaluation module determines the verification result of the target model on the verification image data set. It can be seen that the data development system provided by the present disclosure takes into account the different requirements of different visual tasks for model training in the process of realizing AI development, and realizes more robust model training based on the data enhancement processing method. The entire development process does not require a high level of professional and technical skills, which will significantly improve the development efficiency. At the same time, the entire development system always carries out AI development around a unified visual task, so that the developed model will be more adaptable to the needs of related task scenarios and have higher accuracy.

[0035] Other advantages of the present disclosure will be explained in more detail in conjunction with the following description and drawings.

[0036] It should be understood that the above description is only an overview of the technical solution of the present disclosure, so that the technical means of the present disclosure can be generally understood and then implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present disclosure more obvious and easy to understand, the specific implementation methods of the present disclosure are specifically described below with examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the drawings only illustrate certain embodiments of the present disclosure and therefore should not be regarded as limiting the scope of protection. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings. In the drawings:

[0038] Figure 1 A schematic diagram of a module of a data development system provided by an embodiment of the present disclosure is shown;

[0039] Figure 2 The following is a schematic diagram showing the principle of automatic hyper-parameter search in the data development system provided by the embodiment of the present disclosure;

[0040] Figure 3(a)~3(e) A schematic diagram of a user interface of a data development system provided by an embodiment of the present disclosure at various operation stages is shown. DETAILED DESCRIPTION

[0041] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] In the description of the embodiments of the present disclosure, it should be understood that terms such as "including" or "having" are intended to indicate the presence of the disclosed features, numbers, steps, behaviors, components, parts, or a combination thereof in the specification, and do not exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof.

[0043] Unless otherwise specified, “ / ” means or. For example, A / B can mean A or B. The “and / or” in this article is merely a way to describe the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0044] The terms "first", "second", etc. are used only to distinguish the same or similar technical features for the convenience of description, and should not be understood as indicating or implying the relative importance or quantity of these technical features. Thus, the features defined by "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present disclosure, unless otherwise specified, the term "plurality" means two or more than two.

[0045] Research has found that existing technical solutions usually use professional technical teams to evaluate, write corresponding codes after evaluation, manually debug parameters, and use script training to process AI vision projects. This has the following main shortcomings.

[0046] First, the above method has the problem of high labor cost. Specifically, the evaluators and code writers are generally the front-line technical personnel of the enterprise. The project has a long cycle from evaluation, writing to implementation and deployment, and the comprehensive cost is high. Secondly, the above method is not universal. Specifically, the existing technical solutions are often aimed at a specific scenario such as mask recognition, liveness detection, pedestrian detection, etc. After changing the project scenario, the code needs to be re-adapted and the parameters need to be adjusted. Later project transplantation, maintenance, and iterative upgrades are difficult; thirdly, the above method has the problem of high implementation threshold. Specifically, the overall operation process of the existing solution has high requirements for the development team. Usually, the team developers are composed of people with bachelor's degrees or above, and the final implementation effect of the solution is seriously dependent on the level of the development team personnel; in addition, the above method has the problem of difficulty in transplantation. Specifically, the development platform of the existing solution generally does not have cross-platform characteristics, and the service system can only be used in a single The AI ​​algorithm functions are executed under the operating system (such as Linux system, Windows system or embedded system), and it is difficult for users to migrate across platforms after changing the operating system; finally, the above method has the problem of complex model deployment process. Specifically, since AI algorithms are algorithms that rely heavily on computing resources, most AI development use cases will separate AI intensive computing resources from distributed reasoning deployment resources. For example, computing-intensive supercomputing service centers and embedded devices deployed in intelligent engineering will be physically separated in different geographical locations. For example, intensive AI computing resources are set up in A to train AI models, and a large number of end-side embedded devices are deployed in smart factories such as B. This mode of geographical separation of computing resources will bring difficulties to model training and model deployment. Developers often need to switch between computing-intensive regions and end-side device deployment regions.

[0047] In order to at least partially solve one or more of the above-mentioned problems and other potential problems, the present disclosure provides a data development system to realize AI development for various visual tasks. The whole system has better compatibility and simple operation, which significantly improves the development efficiency.

[0048] See also Figure 1 , which shows a module schematic diagram of a data development system provided by an embodiment of the present disclosure, the system mainly includes:

[0049] An image acquisition module 11, configured to acquire an original image data set in a format matching the target visual task in response to a data acquisition instruction for the target visual task in the user interface;

[0050] A data analysis module 22, configured to respond to a splitting instruction for the original image data set and split the original image data set according to a preset splitting strategy to obtain a split image data set; the split image data set at least includes a training image data set and a verification image data set;

[0051] A model training module 33 is used to respond to a model training instruction for the target visual task, process the training image data set according to a preset data enhancement method, and perform model training based on the enhanced image data set obtained by the processing to obtain a target model adapted to the target visual task;

[0052] The model evaluation module 44 is used to determine the verification result of the target model on the verification image dataset in response to the evaluation instruction for the target model.

[0053] The target visual tasks here include but are not limited to common AI visual tasks such as image classification, target detection, and semantic segmentation. The data development system provided by the disclosed embodiment enables users to complete target visual tasks by interacting with the system platform through the user interface (UI). The overall operation process is convenient and efficient, and large-scale deployment of AI models can be carried out even under low-cost conditions, so that the upstream end can fully and efficiently empower the downstream visual industry with AI.

[0054] The image acquisition module 11 is mainly used to obtain the original image data set in a format that matches the target visual task. The formats of the image data sets supported for different visual tasks are also different. For example, for classification tasks, the system platform supports the import of ImageNet type labeled data sets; for target detection tasks, the system platform supports the import of coco, voc standard data set formats and open source labeling tool labelImg labeled data sets; for semantic segmentation tasks, the system platform supports the import of grayscale type labeled data sets.

[0055] In practical applications, the original image dataset can be constructed based on a preset multi-level dataset construction structure. The multi-level dataset construction structure here can be a construction structure based on different image attribute divisions, such as a three-level construction structure. A part of the image data is extracted at each level so that the constructed original image dataset has a certain degree of randomness, thereby further improving the generalization ability of subsequent model training.

[0056] When it is determined that the task is targeted, the original image dataset matching the target visual task can be obtained based on a pre-configured image format. Here, the dataset matching the format can be obtained directly, or the dataset can be obtained indirectly after image format conversion.

[0057] The data development system provided by the embodiment of the present disclosure can perform data splitting in advance through the data analysis module 22 in order to realize subsequent more automated model training, evaluation and other processes. Specifically, the original image data set is split according to a preset splitting strategy to obtain a split image data set; the split image data set includes at least a training image data set and a verification image data set; that is, the original image data set is split into at least a training image data set and a verification image data set, the former is used for AI model training, and the latter is used for model performance evaluation.

[0058] The data splitting in the disclosed embodiment supports various methods such as default splitting, user-defined splitting, automatic splitting and remote-assisted splitting, and each method can correspond to different splitting strategies. Among them, default splitting: usually the ratio of training set and verification set is divided into 70% and 30% by default; user-defined splitting: supports user-defined setting of training set and verification set splitting ratio; automatic splitting: the system platform will adaptively split the data set according to the number and proportion of each analogy in the user data set to avoid the situation where the number of verification set images is too small or too large; remote-assisted splitting: users can authorize platform experts to access the system platform to assist in splitting and optimize the splitting results.

[0059] In addition, the disclosed embodiment also supports data checking, allowing users to check whether each labeled instance is correctly labeled, whether it contains mixed labels or missing labels, etc.; at the same time, the data check will also count the quantity distribution of each category, so that users can directly see the distribution of each type in the sample (whether there is a long-tail distribution, etc.).

[0060] After sufficient data preparation, model training can be performed based on the model training module 33. Specifically, the training image data set is processed according to a preset data enhancement method, and model training is performed based on the enhanced image data set obtained by the processing to obtain a target model suitable for the target visual task.

[0061] The preset data enhancement methods here may include traditional data enhancement methods, such as image rotation, translation, brightness change, channel change, random erasing and other operations on the data set; it may also include automatic data enhancement methods, that is, the system platform searches for automatic enhancement strategies based on the data set uploaded by the user.

[0062] The data set after data enhancement will be more robust for training, making the trained target model more accurate. In this way, the target model is verified on the verification image data set through the model evaluation module 44, and the evaluation parameters such as precision, recall, F1 score of the target model can be determined through the verification result to evaluate the performance of the target model.

[0063] Among them, the various instructions to which the various modules respond respectively, such as data acquisition instructions for the target visual task, splitting instructions for the original image data set, model training instructions for the target visual task, and evaluation instructions for the target model, etc., can be initiated by the user separately, or can be directly generated under the one-key integrated start button, or can be initiated in other ways. There is no specific restriction here. In actual applications, it can be determined in combination with the actual needs of the user.

[0064] It should be noted that before model training, it is necessary to select a suitable model structure to be trained based on the target visual task. The disclosed embodiment can search for relevant model structures from the model library, such as open source model structures, customized model structures, etc.

[0065] The open source model structure here mainly includes lightweight models and general models. The lightweight model generally has a smaller structure, fewer parameters, and a faster speed, which has a speed advantage on the terminal device. The general model is suitable for a variety of scenarios and generally has higher accuracy, but is slower on the terminal device. If it is to be deployed on the terminal device, it needs to be further compressed.

[0066] The model structure here can be a specific model structure that is only used for a specific image processing chip, such as an imagenet pre-trained model. It has better adaptability to tasks such as image classification, target detection, semantic segmentation, face detection, optical character recognition (OCR), human key point detection, and behavior recognition in specific industry scenarios, and has higher accuracy and convergence speed; the imagenet pre-trained model has better versatility and can adapt to most industries and application scenarios.

[0067] Considering the key role of the model training module 33 in realizing large-scale deployment of AI models in the embodiment of the present disclosure, the following will focus on the model training module 33.

[0068] Considering that traditional data enhancement methods rely on manual parameter adjustment, engineers need to select data enhancement methods for the data set and set initial values ​​for each data enhancement method. This manual data enhancement method has low universality and requires repeated parameter adjustment when changing the data set or task type.

[0069] In order to solve this problem, the data development system provided by the embodiment of the present disclosure includes a built-in automatic data enhancement method, which can be specifically implemented based on the data enhancement unit 331 included in the model training module 33. The data enhancement unit 331 responds to the trigger instruction for the automatic data enhancement strategy under the target visual task, randomly extracts a data enhancement method from the enhancement space, and randomly extracts an enhancement amplitude under the data enhancement method; uses the randomly extracted data enhancement method and its enhancement amplitude to process part or all of the image data in the training image data set to obtain an enhanced image data set, thereby avoiding the problems of low universality and high adjustment complexity of the existing data enhancement that relies on manual parameter adjustment, and has a better enhancement effect.

[0070] In addition, the embodiments of the present disclosure set different data enhancement methods for different task types. For example, for image classification tasks, the system platform supports brightness enhancement, contrast enhancement, saturation enhancement, chroma enhancement, sharpness enhancement, horizontal flipping, vertical flipping, and rotation, a total of 8 data enhancement methods. In actual applications, the enhancement value range corresponding to each data enhancement method is [0.1~1.0], thereby further improving the data enhancement effect under different task types.

[0071] Considering that different hyperparameter settings have a huge impact on the final training effect of the model during model training, this will directly affect the actual performance of the model in industrial deployment scenarios. The data sets provided by AI applications in different industries and fields vary greatly, making it difficult to train the model using a set of initial hyperparameters. In addition, the hyperparameter space formed by hyperparameters is huge, and the parameter adjustment task requires developers to have certain mathematical knowledge. It can be seen that hyperparameter setting has always been a technical problem for model training.

[0072] In order to alleviate the above technical problems, the data development system provided in the embodiment of the present disclosure provides a variety of hyper-parameter setting methods, such as default hyper-parameter setting, user-defined setting, automatic hyper-parameter setting and remote assistance setting, which will be explained in the following aspects.

[0073] First aspect: the model training module 33 in the embodiment of the present disclosure implements automatic super-parameter search by including the first super-parameter search unit 332, specifically:

[0074] The first hyperparameter search unit 332, in response to the first hyperparameter search instruction for the target visual task, selects a target hyperparameter type that matches the target visual task from various hyperparameter types based on the task type of the target visual task; obtains an initialized hyperparameter value corresponding to the target hyperparameter type, and cyclically searches and updates the hyperparameter value based on a preset Bayesian hyperparameter search method until a preset search cutoff condition is reached, thereby obtaining the searched hyperparameter value.

[0075] Among them, based on the task type of the target visual task and the original image dataset acquired under the target visual task, a target hyperparameter type matching the original image dataset is selected from various hyperparameter types.

[0076] In actual applications, users can specify the task type (image classification, object detection, semantic segmentation, etc.) on the system platform, and then upload the dataset under the corresponding task to start the customized hyperparameter search. Figure 2 The automatic hyperparameter search schematic diagram shown in the figure illustrates the entire search process in detail.

[0077] Step 1:

[0078] According to the selected task type and data set location, encapsulate the activation hyperparameters of the corresponding task in the UI interface into auto_parameter.dict:

[0079] {

[0080] "task_type":"Classification",

[0081] "dataset_path":" / root / data / Classification",

[0082] "learning_rate":[0.1,0.01,0.001,0.0001,0.00001],

[0083] "epoch":[10,10000],

[0084] "batch_size":[1,2,4,8,16,32,64,128,256,512,1024,2408],

[0085] "optimizer":["SGD","Adam"],

[0086] "loss":["CrossEntropyLoss","MultiMarginLoss","BCEWithLogitsLoss","FocalLoss"],

[0087] "lr_scheduler":["StepLR","CosineAnnealingLR","LinearLR"]

[0088] }

[0089] Step 2:

[0090] Select the optimizable hyperparameters according to the auto_parameter.dict submitted by the user, activate the Bayesian Hyperopt automatic hyperparameter optimization module, and register the auto_parameter.dict parameters to the hyperparameter optimization module.

[0091] Step 3:

[0092] Set the initial value of the auto_parameter.dict parameter by random selection or user UI interface, use the initial value to iterate, and obtain a new hyperparameter combination based on the posterior distribution result. Repeat this process until the number of iterations is met or the evaluation index is less than the built-in threshold.

[0093] Step 4:

[0094] Return the result dictionary and update the UI front end. At this time, the parameter value of the UI front end is the automatic search result.

[0095] {

[0096] "task_type":"Classification",

[0097] "dataset_path":" / root / data / Classification",

[0098] "learning_rate":0.01,

[0099] "epoch":100

[0100] "batch_size":16,

[0101] "optimizer":"Adam",

[0102] "loss":"CrossEntropyLoss",

[0103] "lr_scheduler":"CosineAnnealingLR"

[0104] }

[0105] It can be seen that through automatic hyperparameter search, the training efficiency can be improved while ensuring the accuracy of model training.

[0106] Second aspect: The model training module 33 in the embodiment of the present disclosure implements automatic super-parameter search by including a second super-parameter search unit 333. Specifically:

[0107] The second super-parameter search unit 333 is used to respond to the user's second super-parameter search instruction for the target visual task and locally obtain the hyper-parameters customized by the user under the second super-parameter search instruction.

[0108] Here, users can customize the hyperparameters according to their needs.

[0109] Third aspect: The model training module 33 in the embodiment of the present disclosure implements automatic super-parameter search by including a third super-parameter search unit 334. Specifically:

[0110] The third super-parameter search unit 334 is used to respond to the expert's third super-parameter search instruction for the target visual task, and remotely obtain the hyper-parameters set with the expert's assistance under the third super-parameter search instruction.

[0111] In actual applications, the system platform provides an external access interface, and platform experts can remotely access customer projects to assist users in configuring appropriate hyperparameters.

[0112] The model training module 33 provided in the embodiment of the present disclosure also includes a model training unit 335, which is used to obtain a target model to be trained for the target visual task, input the processed enhanced image data set into the pre-trained target model for at least one round of model training, and obtain a target model adapted to the target visual task. That is, through one or more rounds of model training, model training can be performed based on the data set, so that the accuracy of the target model obtained by training is also higher.

[0113] The target model to be trained for the target visual task can be selected from candidate models based on the task type of the target visual task. That is, for different task types, the embodiment of the present disclosure can pre-set different training models, and when model training is required for the target visual task, model selection can be performed based on the task type. This ensures the matching degree between the model training and the current visual task, and further enables the model obtained from subsequent training to be better deployed to related visual tasks and be more applicable.

[0114] like Figure 1 As shown, the data development system provided by the embodiment of the present disclosure is also provided with an interface display module 55, which displays the verification result in response to the display request of the target user for the verification result. In a specific application, when it is determined that the verification result needs to be displayed, the user can click the Generate Training Report button to obtain the verification result on the verification set.

[0115] In the disclosed embodiment, if the verification is passed, model reasoning can also be performed, for example, the user can be supported to select a local folder for reasoning, and the interface display module 55 here can obtain at least one image data input by the user, and input the at least one image data into the target model in sequence for model reasoning, determine the reasoning evaluation result, and display the reasoning evaluation result. Here, central processing unit (CPU) reasoning and graphics processing unit (GPU) reasoning can be supported, and the user can obtain the reasoning time, reasoning accuracy, etc. of the target model on different devices as the reasoning evaluation result.

[0116] If the inference accuracy, inference speed, etc. of the model can meet user requirements, then the data development system provided by the embodiment of the present disclosure can also upload the target model to the model cloud store bound to the target user account with one click through the one-click upload module 66, so that the required model can be obtained at any time through the model cloud store.

[0117] In actual applications, users can click the model upload button in the user interface to upload the trained model to the model cloud store with one click. The cloud store opens an exclusive model warehouse for each user. After the model is uploaded, the user can log in to the model cloud store in any network environment to download the model in his own model warehouse.

[0118] In order to further understand the data development system provided by the embodiment of the present disclosure, Figure 3(a)~3(e) The following user interface display effect diagram is used to further illustrate the working principle of this data development system:

[0119] As shown in FIG3(a), the original image dataset corresponding to the current image classification task can be obtained based on the pop-up window guidance of the user interface. The image data in the original image dataset can be extracted based on a pre-built multi-level dataset construction structure. For details, please refer to the above description, which will not be repeated here.

[0120] When the original image data set is obtained, the data can be split, as shown in Figure 3(b), where the training set, validation set, and test set account for 80%, 10%, and 10% respectively. The specific splitting strategy will not be described here.

[0121] Based on each data set obtained by splitting, operations such as model training and verification can be performed. Before the specific model training is performed, the embodiment of the present disclosure is also provided with an automatic data enhancement strategy, as shown in FIG3(c), which is a comparison diagram of the effects before and after enhancement. It can be seen that the target in the image is enhanced, thereby preparing data for the subsequent classification model training task.

[0122] Based on this, the "Start Training" button can be activated to automatically complete the model training process. The verification results such as the learning rate curve can verify the performance of the training model, as shown in Figure 3(d).

[0123] Furthermore, the evaluation entry can be used to perform further reasoning deployment on the trained model to determine the reasoning result, as shown in Figure 3(e).

[0124] It can be seen that the data development system provided by the embodiment of the present disclosure can respond to various operations of the user on the user interface, thereby realizing the full-process AI model deployment, and has stronger applicability, and mainly has the following technical advantages: it can effectively reduce the AI ​​development cost of the enterprise; it can effectively shorten the development cycle of AI projects; it can greatly reduce the iterative maintenance cycle of AI projects; in terms of AI model deployment, it can achieve one-place training and multiple-location deployment, and the deployment process is not constrained by geographical physical controls; the automatic parameter adjustment and remote parameter adjustment functions can greatly reduce the development difficulty of customers, lower the user development threshold, and have stronger applicability.

[0125] In the description of this specification, the description with reference to the terms "some possible embodiments", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure, and the above terms do not necessarily represent the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0126] The embodiments of the present disclosure may also provide methods, devices, equipment and computer-readable storage media that correspond one-to-one with the systems. Therefore, the devices, equipment and computer-readable storage media also have similar beneficial technical effects as the corresponding systems. Since the beneficial technical effects of the systems have been described in detail above, the beneficial technical effects of the devices, equipment and computer-readable storage media will not be repeated here.

[0127] It should be understood by those skilled in the art that the embodiments of the present disclosure can be implemented in the form of methods and apparatus (devices or systems), or computer-readable storage media. Therefore, the present disclosure can be implemented in a complete hardware implementation, a complete software implementation, or an implementation that combines software and hardware. Moreover, the present disclosure can be implemented in the form of a computer-readable storage medium implemented on one or more computer-usable storage media (including but not limited to disk storage, read-only compact disk storage (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0128] The present disclosure is described with reference to the flowchart and / or block diagram of the method, device (equipment or system) and computer-readable storage medium according to the embodiment of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, and the combination of the process and / or box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0129] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, wherein the instruction device implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0131] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0132] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash memory (Flash RAM). The memory is an example of a computer-readable medium.

[0133] Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory, read-only memory, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. In addition, although the operations of the disclosed method are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. In addition, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple sub-steps.

[0134] Although the spirit and principle of the present disclosure have been described above with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A data development system, characterized in that: include: An image acquisition module, configured to acquire an original image data set in a format matching the target visual task in response to a data acquisition instruction for the target visual task in the user interface; A data analysis module, configured to respond to a splitting instruction for the original image data set and split the original image data set according to a preset splitting strategy to obtain a split image data set; the split image data set at least includes a training image data set and a verification image data set; A model training module, for responding to a model training instruction for the target visual task, processing the training image data set according to a preset data enhancement method, and performing model training based on the enhanced image data set obtained by the processing to obtain a target model adapted to the target visual task; The model evaluation module is used to determine the verification result of the target model on the verification image dataset in response to the evaluation instruction for the target model.

2. The data development system according to claim 1, characterized in that: The model training module includes a data enhancement unit; The data enhancement unit is used to randomly extract a data enhancement method from the enhancement space in response to a trigger instruction for the automatic data enhancement strategy under the target visual task, and randomly extract an enhancement amplitude under the data enhancement method; Part or all of the image data in the training image data set is processed using a randomly selected data enhancement method and an enhancement amplitude to obtain an enhanced image data set.

3. The data development system according to claim 1, characterized in that: The model training module includes a first super parameter search unit; The first hyperparameter search unit, in response to the first hyperparameter search instruction for the target visual task, selects a target hyperparameter type that matches the target visual task from various hyperparameter types based on the task type of the target visual task; obtains an initialized hyperparameter value corresponding to the target hyperparameter type, and cyclically searches and updates the hyperparameter value based on a preset Bayesian hyperparameter search method until a preset search cutoff condition is reached, thereby obtaining the searched hyperparameter value.

4. The data development system according to claim 3, characterized in that: The first super-parameter search unit is specifically used to select a target super-parameter type that matches the original image dataset from various super-parameter types based on the task type of the target visual task and the original image dataset obtained under the target visual task.

5. The data development system according to claim 1, characterized in that: The model training module includes a second super parameter search unit; The second super-parameter search unit is used to locally obtain the hyper-parameters customized by the user under the second super-parameter search instruction in response to the user's second super-parameter search instruction for the target visual task.

6. The data development system according to claim 1, characterized in that: The model training module includes a third hyperparameter search unit; The third super-parameter search unit is used to respond to the expert's third super-parameter search instruction for the target visual task, and remotely obtain the hyper-parameters set with the expert's assistance under the third super-parameter search instruction.

7. The data development system according to claim 1, characterized in that: The model training module includes a model training unit; The model training unit is used to obtain a target model to be trained for the target visual task, input the processed enhanced image data set into a pre-trained target model for at least one round of model training, and obtain a target model suitable for the target visual task.

8. The data development system according to claim 7, characterized in that: The model training unit is specifically used for: Obtain multiple candidate models corresponding to each visual task; Determining a candidate model corresponding to the target visual task from the multiple candidate models based on the task type of the target visual task; The determined candidate model is used as a target model to be trained for the target visual task.

9. The data development system according to any one of claims 1 to 8, characterized in that: It also includes an interface display module; The interface display module is used to display the verification result in response to a display request of the target user for the verification result.

10. The data development system according to claim 9, characterized in that: The interface display module is also used for: In response to a target user's reasoning request for the target model, at least one image data input by the user is obtained, and the at least one image data is sequentially input into the target model for model reasoning, a reasoning evaluation result is determined, and the reasoning evaluation result is displayed.

11. The data development system according to claim 10, characterized in that: Also includes: The one-click upload module is used to upload the target model to the model cloud store bound to the target user account with one click when it is determined that the reasoning evaluation result meets the preset reasoning requirements, so that the required model can be obtained at any time through the model cloud store.

12. The data development system according to any one of claims 1 to 8, characterized in that: The image acquisition module is specifically used for: Extracting image data matching the task type of the target visual task according to a preset multi-level data set construction structure; The original image data set is constructed based on the extracted image data.

Images