Model construction method based on training-free neural architecture search and automatic deep learning platform

Abstract

The invention discloses a model construction method based on training-free neural architecture search and an automatic deep learning platform, and relates to the technical field of artificial intelligence. The model construction method based on training-free neural architecture search comprises the following steps: acquiring uploaded data of a user, and labeling the uploaded data; the annotated uploaded data are input into a search module based on an NAS-HPO combined search space; a search module based on the NAS-HPO joint search space simultaneously optimizes the neural architecture parameters and the hyper-parameters to generate one or more candidate neural networks; evaluating one or more candidate neural networks in a training-free manner, and determining a Top-1 network in the candidate neural networks; the Top-1 network is finely tuned; and packaging the Top-1 network after fine tuning into a deployable software package. By adopting the technology provided by the invention, the computing power resource can be effectively saved, the construction speed is accelerated, and the full-process automation is realized to reduce the requirement on the professional degree of a user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a model construction method based on training-free neural architecture search and an automatic deep learning platform applying the same. BACKGROUND

[0002] Auto DL technology aims to reduce the threshold of artificial intelligence application, and supports users to efficiently construct high-performance models without deep AI theoretical background (i.e. non-expert users). At present, the related solutions in the market can be mainly divided into the following two categories: The first category is represented by Google Cloud Auto ML, Baidu PaddlePaddle Easy DL and Aliyun PAI-AutoLearning online hosting Auto DL, which adopts the "training evaluation" strategy, that is, it needs to train each candidate neural network architecture in the search space completely or for a certain number of times. Although this strategy has the advantage of stable performance, it also consumes a lot of computing resources.

[0003] The second category is represented by Huawei Cloud Model Arts, OneClick.AI and other platforms, which are integrated automatic modules based on the first category, but they still cannot break through the traditional training-validation-evaluation closed loop, and their performance is strongly related to the data size. It is not suitable for small and medium-sized data sets (common scenarios for SMEs), and it is easy to cause overfitting due to excessive search. SUMMARY

[0004] The present application provides a model construction method based on training-free neural architecture search and an automatic deep learning platform to solve the problem of large consumption of computing resources in the prior art.

[0005] To solve the above technical problems, the embodiments of the present application provide a model construction method based on training-free neural architecture search, which comprises: Obtaining user uploaded data and labeling the uploaded data; Inputting the labeled uploaded data into a search module based on a NAS-HPO joint search space; The search module based on the NAS-HPO joint search space simultaneously optimizes the neural architecture parameters and the hyperparameters to generate one or more candidate neural networks; Training-free evaluation of one or more candidate neural networks and determination of Top-1 network in the candidate neural network; Fine-tuning the Top-1 network; The fine-tuned Top-1 network is packaged into a deployable software package.

[0006] By means of the search module based on the NAS-HPO joint search space, the structural variables and the hyperparameter variables of the neural architecture parameters are integrated into the same optimization vector, and joint sampling and evaluation are performed in a unified cycle, so that joint search of the network structure and the training hyperparameters is realized, a high-performance matching between the neural architecture parameters and the hyperparameters is obtained, suboptimal solutions caused by traditional step-by-step optimization are avoided, and a final model with superior performance is obtained.

[0007] The performance score is quickly generated by means of the training-free (NUS) technology, and the Top-1 network is fine-tuned, so as to further reduce the cost of computing power demand and speed up the calculation speed. Full-process automation is realized to reduce the demand for user expertise.

[0008] According to another specific embodiment of the present application, the training-free evaluation of the candidate neural network comprises: The performance of the candidate neural network is evaluated and sorted on the CPU server side.

[0009] According to another specific embodiment of the present application, the training-free evaluation of the candidate neural network comprises: Different input samples are determined according to the uploaded data of the user, and different input samples are input into one or more candidate neural networks for one forward propagation, to generate feature mapping in the hidden layer of one or more candidate neural networks; According to the feature mapping and the input sample corresponding to the feature mapping, a linear separability score and / or a feature correlation score of the feature mapping are calculated, to predict the performance score of the corresponding one or more candidate neural networks.

[0010] According to another specific embodiment of the present application, the search module based on the NAS-HPO joint search space simultaneously optimizes the neural architecture parameters and the hyperparameters to generate one or more candidate neural networks, comprising: When the search module jointly optimizes the neural architecture parameters and the hyperparameters, the Bayesian optimization technology is used to optimize the hyperparameters.

[0011] According to another specific embodiment of the present application, the fine-tuning of the Top-1 network comprises: The Top-1 network is fine-tuned on the GPU server side by using the parameter inheritance technology and the transfer learning technology.

[0012] According to another specific embodiment of the present application, the fine-tuning of the Top-1 network further comprises: iteratively fine-tuning the Top-1 network, and calculating a performance score of the Top-1 network in each iteration; terminating the training of the Top-1 network when the performance score of the Top-1 network in the last k iterations is less than a threshold value and / or, obtaining test data, and testing the iteratively fine-tuned Top-1 network.

[0013] According to another specific embodiment of the present application, the model construction method based on the training-free neural architecture search further comprises performing Git-like version management on at least one of the dataset version, the model version, and the deployment configuration file version, to support user rollback operation on any historical version.

[0014] According to another specific embodiment of the present application, the obtaining of the uploaded data of the user and the labeling of the uploaded data comprise: invoking a pre-trained model using a semi-automatic labeling engine to label the uploaded data; and / or, manually adjusting the labeling of the uploaded data, or manually labeling the uploaded data.

[0015] According to another specific embodiment of the present application, the encapsulation of the fine-tuned Top-1 network into a deployable software package comprises: generating at least one of a RESTful application programming interface file, a software development kit (SDK), and a Docker image file corresponding to the Top-1 network; wherein the deployable software package can support a local server or a cloud environment.

[0016] The embodiments of the present application also provide an automatic deep learning platform applying the above-described model construction method based on the training-free neural architecture search, which comprises from bottom to top in sequence: an IaaS layer comprising a CPU server and / or a GPU server, the CPU server being used to evaluate the score of the candidate neural network in a training-free manner, and / or the GPU server being used to fine-tune the Top-1 network; a PaaS layer integrated with a data automatic labeling engine, an automatic feature engineering engine, a pre-trained model automatic selection engine, a neural architecture automatic search engine, a hyperparameter automatic optimization engine, and a model SDK automatic encapsulation engine; a SaaS layer used to provide one or more application services in image classification, object detection, image segmentation, and natural language processing By adopting the technical scheme, the heterogeneous hybrid computing mode of the CPU and the GPU is used to realize that the CPU server bears the computing load of the training-free neural architecture search, and is matched with the CPU parallel computing characteristics to realize fast evaluation of the CPU, and the GPU fine-tuning is used to reduce the computing power consumption. In addition, based on the top-down SaaS application traction, the automatic driving of the core PaaS engine chain, and the heterogeneous cooperation of the underlying IaaS resources, the standardization and automation of AI development are realized, and the "no code" development AI is realized.

[0017] Compared with the traditional development mode, at least the steps of data labeling (semi-automatic labeling of the data set in the application), model configuration (training-free NAS and GPU fine-tuning based on the application), and deployment configuration (one-key deployment of the application) are saved, and the whole process operation of the platform is realized. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The schematic embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a system architecture schematic diagram of an automatic deep learning platform of an embodiment of the present application; Figure 2 is a partial architecture schematic diagram of the PaaS layer in the application; Figure 1 Figure 3 is a module schematic diagram of an automatic deep learning platform of an embodiment of the present application; Figure 4 is a full life cycle management schematic diagram of an AI model of an automatic deep learning platform of an embodiment of the present application; Figure 5 is a flow schematic diagram of a model construction method based on training-free neural architecture search of an embodiment of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application are described below by specific specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the description. Although the description of the present application will be introduced in combination with the preferred embodiments, this does not mean that the features of the present application are limited to the embodiments. On the contrary, the purpose of introducing the present application in combination with the embodiments is to cover other options or modifications that can be extended based on the claims of the present application. In order to provide a deep understanding of the present application, many specific details will be included in the following description. The present application can also be implemented without using these details. In addition, in order to avoid confusion or obscure the focus of the present application, some specific details will be omitted in the description. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.​

[0020] In order to make the objects, technical solutions and advantages of the present application clearer, the following will explain the content of the present application in conjunction with the drawings.

[0021] The prior art highly relies on the "training-evaluation" cycle, which results in high economic and time costs for searching a network structure even requiring thousands of GPU hours. Therefore, in order to reduce the computing resources, referring to FIG. 1, the present application provides an automatic deep learning platform, which comprises from bottom to top: Figures 1 to 3 The IaaS layer comprises CPU servers and / or GPU servers, the CPU servers are used to evaluate the scores of candidate neural networks without training, and / or the GPU servers are used to fine-tune a Top-1 network.

[0022] The PaaS layer is integrated with a data automatic labeling engine, an automatic feature engineering engine, a pre-trained model automatic selection engine, a neural architecture automatic search engine, a hyperparameter automatic optimization engine and a model SDK automatic packaging engine. The SaaS layer is used to provide one or more application services in image classification, target detection, image segmentation and natural language processing.

[0023] By using the above technical solution, the IaaS layer resources are used to realize the cooperation with the PaaS layer algorithms to efficiently create AI applications. Specifically, at the IaaS layer, the CPU server end performs "evaluation without training" to convert the model search process requiring a large amount of GPU resources into a CPU computing task, thereby significantly reducing the computing consumption.

[0024] Alternatively, a Top-1 network is fine-tuned based on the GPU server end to further reduce the computing consumption. The Top-1 network represents the best architecture ranked first among all candidate neural networks. For example, the present application only retains one Top-1 network as the optimal solution to quickly promote subsequent deployment tasks, which has a lower computing cost compared with retaining multiple candidate neural networks.

[0025] In addition, the search-tuning cycle in the conventional technology is long, which generally takes 1 to 3 days, and in special cases, it can take up to 7 days. In the present application, in order to speed up the search-tuning cycle, the PaaS layer is used to place the neural architecture search (NAS) and the parameter optimization in the same joint search space. For example, Figure 2 ​As shown in the diagram, Neural Architecture Search (NAS) and Parameter Optimization are used to automate and optimize the entire model development process to meet the user's (target enterprise's) need for "rapid delivery".

[0026] Specifically, Neural Architecture Search (NAS) is responsible for finding the optimal network architecture, while Parameter Optimization (PO) is responsible for training the best model within that architecture. Compared to manually setting the search space and resource budget, it effectively reduces the requirements for the size of user data, avoids overfitting on small to medium-sized datasets, and improves search efficiency.

[0027] In one specific implementation, the parameter optimization shown in the figure is extended to hyperparameter optimization (HPO) to achieve a broader optimization process, which includes not only the optimization of model weights, but also the optimization of hyperparameters such as learning rate and batch size.

[0028] Among them, such as Figure 2 As shown, the engines integrated into the PaaS layer satisfy the complete automated engine chain from user-uploaded data to final model deployment.

[0029] As the top layer, the SaaS layer provides various AI application services to meet users' needs for immediate use. That is, after users upload data, the platform automatically completes the deployment process, and the SaaS layer adapts the neural network model to the user's required application. It is worth noting that the neural network model mentioned in this application's technical solution can be a neural network model used in tasks such as image recognition, object detection, image segmentation, and natural language processing. The type of neural network model is not limited to any neural network model such as RNN, GRU, LSTM, conditional generative models, reinforcement learning models, or Transformer (a model based on a multi-head attention mechanism).

[0030] See Figures 4 to 5 As shown, this application also provides a model construction method based on training-free neural architecture search, which is applied to the aforementioned automated deep learning platform. The model construction method based on training-free neural architecture search includes: Step S1: Obtain the user's uploaded data and label the uploaded data.

[0031] In the data processing stage, the user uploads the unlabeled data to the data center in the data automatic labeling engine of the PaaS layer of the platform, wherein, when the user's goal is an image classification task, the unlabeled data is image data; when the user's goal is a text classification task, the unlabeled data is text data, and the data is usually 1 group, which can be adaptively adjusted according to actual needs, and the application does not limit this. The front end of the platform generates corresponding data entries according to the data set name, task type, uploaded data and the like input by the user, and stores them in the database.

[0032] The user can start the labeling program for the uploaded data by clicking "automatic labeling". The labeling program for the uploaded data can use a semi-automatic labeling engine to call a pre-trained model to label the uploaded data. Specifically, the platform background task management process starts the preset corresponding task pre-trained model, predicts the data in the data set one by one, and the prediction result is stored as a pseudo label in the corresponding database.

[0033] The preset corresponding task pre-trained model includes but is not limited to Resnet, MobileNet-V2, Yolo5, R2PlusD_18, PanopticDeepLab, Deepsort+Yolo5, BERT, etc. Among them, the user can query whether the labeling task is completed by checking the data set state.

[0034] In a possible application scenario, the user can click "manual labeling" to view the automatically labeled data set, and can adjust the data labeling information through "manual labeling".

[0035] For example, in a target tracking task, the user organizes an mp4 video file, compresses it and uploads it to the data center in the data automatic labeling engine of the PaaS layer of the platform. The platform can use a YOLO pre-trained model to perform target detection on the first frame of image, and based on the detection result, a built-in tracking algorithm of an OpenCV pre-trained model is used for tracking labeling. Exemplarily, in the "manual labeling" stage, the user can modify the labeling information by playing the video through the front end and displaying the labeling box, or by dragging the edge of the labeling box with a mouse.

[0036] In another possible application scenario, the user can also abandon "automatic labeling" and use "manual labeling" for all selected data. This ensures the adaptability of the subsequent model building.

[0037] Step S2, input the labeled uploaded data into the search module based on the NAS-HPO joint search space; so that the platform can determine and evaluate the candidate architecture based on the labeled data.

[0038] Specifically, the search module based on the NAS-HPO joint search space represents a unified optimization unit with the ability of (differentiable) architecture sampling and hyperparameter sampling. It integrates the structure variables of neural architecture parameters and hyperparameter variables into the same optimization vector and performs joint sampling and evaluation in a unified cycle to realize the joint search of network structure and training hyperparameters. For example, the search module instantiates the corresponding sub-neural network and hyperparameter configuration during forward propagation, and updates the network weights through gradient during backward propagation to realize the following step S3.

[0039] Step S3, the search module simultaneously optimizes the neural architecture parameters and the hyperparameters to generate one or more candidate neural networks. Compared with the traditional "search architecture first and then adjust parameters" method, step S3 can find the global optimal matching between architecture and hyperparameters in the search stage. In other words, the traditional technicians always focus on improving the search speed and accelerating the parameter adjustment rate, and in order to ensure the optimal selection of subsequent optimal parameters and architecture, the number of candidate neural networks that need to be reserved and optimized cannot be too small.

[0040] In this application, the simultaneous optimization of neural architecture parameters and hyperparameters by the search module substantially expands the search range to network depth, width, operation type, learning rate, batch size, etc., and maximally reduces human intervention, which can reduce the demand for candidate neural networks in the early stage, and also reduce the demand for parameter adjustment computing power in the later stage, greatly improving the search matching degree and search efficiency.

[0041] In a specific application scenario, a user can create a model through the "model center" of the above-mentioned platform, so that the platform completes model creation and generates a model entry according to the task type selected by the user, and stores it in the database. When the user clicks "model training" on the platform and selects the target labeled data set, the above-mentioned platform will automatically start model training.

[0042] Specifically, the above-mentioned platform based on the NAS-HPO joint search space first uses a pre-training NAS to efficiently search a neural architecture suitable for the target data set, and simultaneously searches the hyperparameters according to the current optimal neural architecture searched and the target data set (such as a user-uploaded data set or a user-selected historical uploaded data set).

[0043] For example, when the current optimal neural architecture searched is optimized for hyperparameters, a Bayesian optimization technique is used.

[0044] In a specific application scenario, the above-mentioned platform starts model training without human intervention throughout the whole process. In addition, in another specific application scenario, a user can view training information in model training, including but not limited to viewing loss / accuracy curves, training progress, etc.

[0045] Step S4, evaluating one or more candidate neural networks and determining a Top-1 network in the candidate neural networks; The NAS-HPO joint search space described above uses the NAS-Benchmark-ALL open source framework as a benchmark for evaluating the backend neural architecture search algorithm to ensure the effectiveness of the platform core algorithm. For example, based on the unified NAS-Benchmark-ALL open source framework, the same NAS-HPO joint search space and the same target data set are used to test different indicators to calculate the ranking correlation score.

[0046] Meanwhile, in a possible application scenario, if the feasibility of replacing the existing search algorithm described above is subsequently evaluated, the above evaluation benchmark can be used to quantitatively determine the advantages and disadvantages between the existing search algorithm and the target new algorithm, thereby assisting the backend developer in determining the replacement feasibility.

[0047] Specifically, in a specific embodiment, the step S4 includes determining different input samples according to the user's uploaded data, and inputting the different input samples into one or more candidate neural networks for a forward propagation to generate feature mappings in the hidden layers of the one or more candidate neural networks.

[0048] According to the feature mappings and the input samples corresponding to the feature mappings, a linear separability score and / or a feature correlation score of the feature mappings are calculated to predict the performance score of the corresponding one or more candidate neural networks.

[0049] By calculating the "linear separability score" and the "feature correlation score" to estimate the network performance, there is no need for back propagation. That is, in the embodiments of the present application, the network performance is evaluated only according to the feature mappings generated in the hidden layers of the one or more candidate neural networks by different input samples of the initial network. Thus, the NAS is greatly accelerated without significant performance degradation, and the dependence on computing resources in the NAS process is reduced. That is, the platform described above can quickly evaluate one or more candidate neural networks using only the CPU end.

[0050] Specifically, the linear separability score is used to measure the ability of the neural network to distinguish a set of features into different categories, which can be achieved by quantifying the "inter-class separation degree". For example, the Fisher discriminant rate, the ratio of inter-class dispersion to intra-class dispersion of linear discriminant analysis (LDA), etc.

[0051] The feature correlation score can measure the correlation of non-linear features compared to the linear separability score, to quantify the relevance of a certain feature to the prediction target, such as the Spearman rank correlation coefficient, mutual information, etc.

[0052] Step S5, fine-tuning the Top-1 network.

[0053] Since fine-tuning training usually requires a large number of matrix operations, the computing task is placed on the GPU server side for execution. The parallel computing architecture of the GPU makes the GPU more suitable for such computing than the CPU, that is, distributed execution (the search phase is completed on the CPU, and the training phase is completed by the GPU) can ensure the speed of fine-tuning.

[0054] Further, the GPU server side uses parameter inheritance technology and transfer learning technology to fine-tune the Top-1 network. Among them, the parameter inheritance technology means that the above-mentioned Top-1 network is started from a pre-trained model weight, for example, the weight of the ResNet model pre-trained on ImageNet. That is, the Top-1 network has the ability to extract general image features (such as edges, textures, shapes) in the initial state. That is, the present application is based on parameter inheritance technology rather than training from scratch, and the pre-trained weight can provide regularization constraints to adapt to the case of small data volume and prevent overfitting.

[0055] Among them, the transfer learning technology can be based on Figure 3 The pre-stored "model center" realizes the migration of knowledge and model structure in the task to the target model. Specifically, if the user's goal is an image classification task, the target field is image classification, which can migrate the knowledge and model structure learned from the source field (such as ImageNet classification) to the above-mentioned image classification task. Thus, based on the new data uploaded by the current user, the model training with a shorter time and a smaller learning rate can be realized. Further reduce the demand for data volume and prevent overfitting.

[0056] Step S50, obtaining test data, testing the iteratively fine-tuned Top-1 network.

[0057] After the model is fine-tuned, the user can upload one or a group of data again, or use the current target data set to evaluate the trained model. Illustratively, the user clicks "model test" on the above-mentioned platform, then batch tests or individually tests the model, and uploads data or selects a data set, which can execute the prediction task. In an application scenario, the user can view the prediction result and export the prediction result.

[0058] Step S6, encapsulating the fine-tuned Top-1 network into a deployable software package.

[0059] Exemplarily, at least one of a RESTful application programming interface file, a software development kit (SDK), and a Docker image file corresponding to the Top-1 network is generated at the deployment center of the platform. That is, the user can generate the RESTful API / SDK with one click, and the Docker image can be deployed locally or in the cloud. Thus, the platform can realize the automation of the AI application "end to end".

[0060] In a specific embodiment, the step S5 comprises: iteratively fine-tuning the Top-1 network, and calculating the performance score of the Top-1 network in each iteration; and exemplarily, the score of each model is evaluated after each iteration.

[0061] When the performance score of the Top-1 network in the last k iterations is improved by less than a threshold value ε, the training of the Top-1 network is terminated. Exemplarily, the performance score can be an accuracy, an F1 score, a loss value, or the like, which is not limited in the present application.

[0062] In a specific embodiment, k = 5 and ε = 0.5%, that is, when the performance score is improved by less than 0.5% in the last 5 iterations, the search is automatically terminated to obtain the best balance between the adaptive model and the computing power consumption.

[0063] In a possible embodiment, the model construction method based on the training-free neural architecture search further comprises: performing Git-like version management on at least one of a data set version, a model version, and a deployment configuration file version, so as to support the user to perform a rollback operation on any historical version.

[0064] By using the above technical solution, through version management: the data set, the model, and the deployment configuration all support Git-like version rollback, so as to ensure the reproducibility and reliability of each step in the entire AI life cycle. Meanwhile, when the iteration process does not meet the user's expectation, such as when the latest deployed model (version v2) fails or degrades in performance subsequently, the user can roll back to the previous stable and reliable version v1, thereby quickly recovering the service and ensuring the business continuity.

[0065] In another possible embodiment, the step S6 further comprises local / private cloud deployment, that is, deployment in a user's own machine room or a user-exclusive private cloud environment, so that the related data does not go out of the enterprise intranet, and meets the requirements of national security and data protection.

[0066] Further, when the local / private cloud software and hardware are delivered, a CPU server or a GPU server can be selected.

[0067] The above detailed description merely illustrates preferred and possible non-limiting implementations of the application, and is not intended to limit the scope of the application. Other equally effective embodiments and modifications will be apparent to those skilled in the art, and are intended to be included within the scope of the application.

Claims

1. A model construction method based on a training-free neural architecture search, characterized in that, The method comprises the following steps: obtaining user upload data and labeling the upload data; inputting the labeled upload data into a search module based on a NAS-HPO joint search space; simultaneously optimizing neural architecture parameters and hyperparameters based on the search module of the NAS-HPO joint search space to generate one or more candidate neural networks; training-free evaluation of one or more candidate neural networks and determination of a Top-1 network in the candidate neural networks; fine-tuning the Top-1 network; packaging the fine-tuned Top-1 network into a deployable software package. 2.The model construction method based on the training-free neural architecture search according to claim 1, wherein, The training-free evaluation of the candidate neural networks comprises: performance evaluation and sorting of the candidate neural networks on a CPU server side. 3.The model construction method based on the training-free neural architecture search according to claim 1, wherein, The training-free evaluation of the candidate neural networks comprises: determining different input samples according to the upload data of the user, and inputting different input samples into one or more candidate neural networks for a forward propagation, to generate feature maps in the hidden layers of one or more candidate neural networks; calculating the linear separability score and / or feature correlation score of the feature maps according to the feature maps and the input samples corresponding to the feature maps, to predict the performance score of the corresponding one or more candidate neural networks. 4.The model construction method based on the training-free neural architecture search according to claim 1, wherein, The search module based on the NAS-HPO joint search space simultaneously optimizes neural architecture parameters and hyperparameters to generate one or more candidate neural networks, comprising: When the search module jointly optimizes neural architecture parameters and hyperparameters, a Bayesian optimization technique is used to optimize hyperparameters. 5.The model construction method based on the training-free neural architecture search according to claim 1, wherein, The fine-tuning of the Top-1 network comprises: using parameter inheritance technology and transfer learning technology to fine-tune the Top-1 network on a GPU server side. 6.The model construction method based on the training-free neural architecture search according to claim 5, wherein, The fine-tuning of the Top-1 network further comprises: iterative fine-tuning of the Top-1 network and calculation of the performance score of the Top-1 network at each iteration; when the performance score of the Top-1 network improves by less than a threshold value ε for k consecutive iterations, terminate training of the Top-1 network and / or, obtaining test data and testing the iteratively fine-tuned Top-1 network.

7. The model construction method based on the training-free neural architecture search according to any one of claims 1 to 6, characterized in that, The model construction method based on training-free neural architecture search further comprises Git-like version management of at least one of the dataset version, the model version, and the deployment configuration file version, to support user rollback operation on any historical version.

8. The model building method based on the training-free neural architecture search according to any one of claims 1 to 6, characterized in that, The method of obtaining user upload data and labeling the upload data comprises: using a semi-automatic labeling engine to call a pre-trained model to label the upload data; and / or, manually adjusting the labeling of the upload data or manually labeling the upload data. 9.The model construction method based on the training-free neural architecture search according to claim 1 or 8, characterized in that, The method of packaging the fine-tuned Top-1 network into a deployable software package comprises: generating at least one of a REST ful application programming interface file, a software development kit SDK, and a Docker image file corresponding to the Top-1 network; wherein the deployable software package can support local servers or cloud environments.

10. An automated deep learning platform, characterized in that, The model construction method based on the training-free neural architecture search according to any one of claims 1 to 9, wherein the automatic deep learning platform comprises from bottom to top in sequence: an IaaS layer comprising CPU servers and / or GPU servers, the CPU servers being used to evaluate scores of the candidate neural networks in a training-free manner, and / or the GPU servers being used to fine-tune the Top-1 network; a PaaS layer integrated with a data automatic labeling engine, an automatic feature engineering engine, a pre-trained model automatic selection engine, a neural architecture automatic search engine, a hyperparameter automatic optimization engine, and a model SDK automatic packaging engine; a SaaS layer used to provide one or more application services in image classification, object detection, image segmentation, and natural language processing.

Images