System and Method for Recommending Deep Learning Framework based on Feature Selection

Abstract

The present invention relates to a feature selection-based deep learning framework recommendation system and method that can receive user specifications as input and recommend an optimal deep learning framework for implementing a user's algorithm, thereby enabling efficient application of an optimal deep learning framework suitable for a learning target. The system comprises: a framework attribute / feature input unit that inputs a model template and framework attribute information for framework selection; a user specification input unit that inputs an input deep learning model to be converted and data attribute information to be used; a deep learning framework recommendation unit composed of a model data analyzer and a framework selector that receives framework attributes / features and user specifications and recommends a deep learning framework based on feature selection; and a recommendation framework information output unit that outputs recommendation framework information and a framework optimization model design most suitable for the user's algorithm.

Description

Technology Field

[0001] The present invention relates to the selection of a deep learning framework, and specifically to a feature selection-based deep learning framework recommendation system and method that receives user specifications as input and recommends an optimal deep learning framework for implementing the user's algorithm, thereby enabling the efficient application of an optimal deep learning framework suitable for a learning target. Background Technology

[0002] Deep learning refers to performing machine learning using an Artificial Neural Network (ANN) with multiple layers, and is also called deep learning. Therefore, deep learning is not a completely different concept from machine learning, but can be considered a type of machine learning.

[0003] In machine learning, humans had to manually analyze and decide which features to extract from the various features of the data to be learned, whereas in deep learning, the machine automatically extracts features from the data to learn. As such, the biggest difference between deep learning and machine learning can be seen as the machine's ability to self-learn.

[0004] Therefore, deep learning can be defined as a technology in which machines automatically learn important patterns and rules from large-scale data and perform decision-making or prediction based on them.

[0005] Numerous algorithms of various forms, such as DNN, CNN, RNN, RBM, and DBN, have been developed and utilized in artificial neural network algorithms used in deep learning, and it has become common to combine two or more algorithms to solve a single problem.

[0006] Having to re-implement an already proven algorithm every time it is used is a very inefficient approach.

[0007] Deep learning frameworks enable developers to use them quickly and easily by providing numerous already verified libraries and various deep learning algorithms that have been pre-trained.

[0008] This frees developers from the wasteful task of implementing redundant functions, allowing them to focus solely on developing core algorithms for problem solving.

[0009] Deep learning, built using neural networks, is inspired by the structure and function of the brain. Deep learning is a tool that enables innovative advancements in AI.

[0010] Most deep learning frameworks, such as PyTorch, TensorFlow, and ONNX, use directed acyclic graphs to represent neural networks.

[0011] To use deep learning effectively, algorithms for model functions such as predicting and identifying objects must be defined, and a large amount of data is required to train the model for better accuracy.

[0012] There are many frameworks for such deep learning, and each framework has a different purpose.

[0013] TensorFlow is an open-source platform developed by Google. It is JavaScript-based and provides various tools for building and deploying models in the browser.

[0014] The built model can be deployed to mobile and embedded devices and supports large-scale production environments.

[0015] PyTorch is an open-source platform based on the Torch library developed by Facebook. It aims to accelerate the entire process from research prototyping to production deployment and is suitable for applications such as natural language processing and computer vision.

[0016] ONNX is a project jointly developed by Microsoft and Facebook. It is a system designed for the development of machine learning and deep learning models, providing access to hardware optimization and libraries for optimizing overall hardware system performance.

[0017] Keras is an open-source framework that runs on Tensorflow, Theano, Microsoft Cognitive Toolkit, and PlaidML. It provides data parallelism to offer high speed and facilitates the acceleration of model data training time.

[0018] Sonnet was developed by DeepMind, built on TensorFlow, and provides an advanced library for building complex neural network structures and develops and creates basic Python objects corresponding to specific parts of the neural network.

[0019] MXNET is an open source framework that enables fast model training and offers excellent scalability. It supports multiple programming languages ​​such as C++, Python, Julia, Matlab, and JavaScript, and supports multiple GPUs with fast context switching.

[0020] Gluon is an open-source framework that provides a concise API for defining machine learning and deep learning models by utilizing optimized neural network components.

[0021] It is suitable for the development and testing of neural network models, and is suitable for developers new to machine learning.

[0022] In addition to these, there are various deep learning frameworks, and each deep learning framework has different advantages. Among these, there is a need to select and present the optimal deep learning framework to the user according to their purpose.

[0023] With deep learning frameworks, you can obtain good results by selecting the appropriate framework depending on the target you intend to analyze or predict.

[0024] Therefore, there is a need for the development of useful technology for selecting and applying the optimal deep learning framework suitable for the desired learning target. Prior art literature

[0025] Republic of Korea Published Patent No. 10-2021-0123110 Republic of Korea Registered Patent No. 10-2257028 The problem to be solved

[0026] The present invention aims to solve the problems of the deep learning framework selection technology of the prior art and to provide a feature selection-based deep learning framework recommendation system and method that can recommend an optimal deep learning framework for implementing the user's algorithm by receiving user specifications as input, thereby enabling the efficient application of the optimal deep learning framework suitable for the learning target.

[0027] The present invention aims to provide a feature selection-based deep learning framework recommendation system and method that enables a user to make a better choice when selecting a deep learning framework, thereby allowing the user to select the correct framework, save time in implementing algorithms by using the framework's built-in functions to easily build algorithms, and verify whether the recommended framework can be used in the target environment.

[0028] Other objectives of the present invention are not limited to those mentioned above, and other unmentioned objectives will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0029] A feature selection-based deep learning framework recommendation system according to the present invention for achieving the above-mentioned purpose comprises: a framework attribute / feature input unit for inputting a model template and framework attribute information for framework selection; a user specification input unit for inputting an input deep learning model to be converted and data attribute information to be used; a deep learning framework recommendation unit composed of a model data analyzer and a framework selector, which receives the framework attribute / feature and user specification and recommends a feature selection-based deep learning framework; and a recommendation framework information output unit that outputs recommendation framework information and a framework optimization model design most suitable for the user's algorithm.

[0030] Here, the user specification code used as a deep learning input for the framework recommendation system is characterized by including information such as the project's purpose, backend environment, training data, and the user's algorithm written in pseudocode.

[0031] A deep learning framework recommendation method based on feature selection according to the present invention for achieving other purposes is characterized by comprising: a user specification input step receiving a user specification as input; a model data analysis step performing grammar analysis and data input analysis using a model data analyzer; a deep learning framework recommendation step performing algorithm extraction, feature selection, and framework scoring using a framework selection algorithm to recommend a deep learning framework based on feature selection; and a recommendation framework information output step outputting a recommended deep learning framework and a framework optimization model design most suitable for the user's algorithm. Effects of the invention

[0032] The feature selection-based deep learning framework recommendation system and method according to the present invention, as described above, implement the following effects.

[0033] First, by receiving user specifications as input and recommending the optimal deep learning framework for implementing the user's algorithm, the application of the optimal deep learning framework suitable for the training target is efficiently carried out.

[0034] Second, by providing users with the ability to make a better choice when selecting a deep learning framework, users can use the framework's built-in functions to easily build algorithms, thereby saving time in implementing algorithms and verifying whether the recommended framework can be used in the target environment. Brief explanation of the drawing

[0035] FIG. 1 is a configuration diagram of a feature selection-based deep learning framework recommendation system according to the present invention. FIG. 2 is a configuration diagram showing an example of a user specification code for a text file format of a user-defined format. FIG. 3 is a flowchart illustrating a feature selection-based deep learning framework recommendation method according to the present invention. FIG. 4 is a flowchart illustrating the framework scoring algorithm process according to the present invention Specific details for implementing the invention

[0036] Hereinafter, preferred embodiments of the feature selection-based deep learning framework recommendation system and method according to the present invention will be described in detail as follows.

[0037] The features and advantages of the feature selection-based deep learning framework recommendation system and method according to the present invention will become apparent through the detailed description of each embodiment below.

[0038] FIG. 1 is a configuration diagram of a feature selection-based deep learning framework recommendation system according to the present invention, and FIG. 2 is a configuration diagram showing an example of a user specification code in a text file format of a user-defined format.

[0039] The feature selection-based deep learning framework recommendation system and method according to the present invention receives user specifications as input and can recommend an optimal deep learning framework for implementing the user's algorithm, thereby enabling the efficient application of an optimal deep learning framework suitable for the learning target.

[0040] The present invention enables users to make a better choice when selecting a deep learning framework, and by selecting the correct framework, users can use the framework's built-in functions to easily build algorithms, thereby saving time in implementing algorithms and verifying whether the recommended framework can be used in the target environment.

[0041] As shown in FIG. 1, the feature selection-based deep learning framework recommendation system according to the present invention includes a framework attribute / feature input unit (10) for inputting a model template and framework attribute information for framework selection, a user specification input unit (20) for inputting an input deep learning model to be converted and data attribute information to be used, a deep learning framework recommendation unit (30) composed of a model data analyzer and a framework selector that receives the framework attribute / feature and user specification and recommends a deep learning framework based on feature selection, and a recommendation framework information output unit (40) that outputs the recommendation framework information most suitable for the user's algorithm and a framework optimization model design.

[0042] Figure 2 shows an example of a user specification code in the form of a text file of a user-defined format. The user specification code is a deep learning input for a framework recommendation system and includes information such as the project's objective, backend environment, training data, and the user's algorithm written in pseudocode.

[0043] The feature selection-based deep learning framework recommendation method according to the present invention is specifically described as follows.

[0044] FIG. 3 is a flowchart illustrating a deep learning framework recommendation method based on feature selection according to the present invention.

[0045] The deep learning framework recommendation method based on feature selection according to the present invention comprises a user specification input step (S301) that receives a user specification as input, a model data analysis step (S302) that performs grammar analysis and data input analysis using a model data analyzer, a deep learning framework recommendation step (S303) that recommends a deep learning framework based on feature selection by algorithm extraction, feature selection, and framework scoring using a framework selection algorithm, and a recommendation framework information output step (S304) that outputs a recommended deep learning framework and a framework optimization model design that is most suitable for the user's algorithm.

[0046] The model data analysis steps are specifically explained as follows.

[0047] The model data analysis step involves analyzing the grammar of the user specification and data input through a grammar analyzer and a data input analyzer.

[0048] At this stage, the grammar analyzer checks for grammatical errors in the user specification.

[0049] For example, if information other than that predefined by the system is entered as input to the frontend target device, a system error occurs.

[0050] If there are no grammatical errors, the user specification entered as input for model data analysis is output without causing system errors.

[0051] The data input analyzer performs data validation using the CSV file path as input; if invalid data is found, it generates a system error, and if no errors are found, the system assigns the input data to the appropriate variable without generating an error.

[0052] The specific steps of the deep learning framework recommendation process are as follows.

[0053] The framework selector consists of algorithm extraction, feature selection, and framework scoring, and algorithm extraction extracts user algorithms by utilizing the procedure analyzer and function analyzer.

[0054] Feature selection obtains features of the input data by analyzing the relationship between the extracted user algorithm and the input data.

[0055] Framework scoring assigns a score to the most suitable deep learning framework based on selected features.

[0056] In the recommendation framework information output step, information about the recommended deep learning framework and the framework optimization model design are output based on the assigned score.

[0057] Figure 4 is a flowchart illustrating the framework scoring algorithm process according to the present invention.

[0058] The system framework and user specifications are described as follows.

[0059] The system proposed in this invention has information on various deep learning frameworks predefined. It performs matching with user specification information received as input, returns a list of supported frameworks, and then recommends the framework with the highest score.

[0060] Information that may be included in the user specification includes 'program purpose', 'device on which deep learning is to be performed', 'model size', 'data type', etc.

[0061] Feature selection and user specification analysis consist of framework information analysis, user algorithm analysis, and input data feature extraction.

[0062] The selected features of the input data and the framework scoring algorithm utilizing them are explained as follows.

[0063] A framework scoring algorithm is essential to recommend the optimal deep learning framework based on selected features of the user specification.

[0064] The system starts by patching various predefined deep learning framework information, and then patches the features of the user specification obtained through the 'feature selection' process.

[0065] If the feature value of a feature of a user specification exists in the feature value of a feature of a currently selected deep learning framework, 'score(feature value) * weight(feature)' is performed for each feature defined in the deep learning framework to generate the score of the corresponding feature.

[0066] This is performed for each feature of the user specification. As a result, the scores of all features are summed to obtain the score of the corresponding framework.

[0067] Through this, if all features of the user specification are supported by the selected deep learning framework, the framework is added to the list of supported frameworks, and if the selected deep learning framework does not support the feature values ​​of the features included in the user specification, another deep learning framework existing in the system is selected and the same process is performed.

[0068] When the above process is performed for all deep learning frameworks existing in the system, a list of supported frameworks is returned, and the system recommends the deep learning framework with the highest score among them.

[0069] For example, in the case of 'Device' among the various features of the user specification, the feature of the user specification is assumed to be 'Device' and the feature value is 'Android'.

[0070] Assuming the system framework is TensorFlow, first check if there is a 'device' feature among TensorFlow's features; if not, select another deep learning framework, and if the 'device' feature exists in the framework, there are various feature values ​​(CUDA, TPU, CPU, Android, etc.).

[0071] For each feature value, it is checked whether it is identical to the feature in the user specification; if it is identical, the score of the feature value is obtained, and the obtained score is multiplied by the feature's weight to perform scoring for that feature.

[0072] This applies equally to all other features, and framework scoring is performed by summing the scores for all features.

[0073] The feature selection-based deep learning framework recommendation system and method according to the present invention described above receive user specifications as input and recommend an optimal deep learning framework for implementing the user's algorithm, thereby enabling the efficient application of an optimal deep learning framework suitable for the learning target.

[0074] As explained above, it will be understood that the present invention is implemented in a modified form without departing from the essential characteristics of the invention.

[0075] Therefore, the described embodiments should be considered in an illustrative rather than a limiting sense, and the scope of the invention is defined by the claims rather than the foregoing description, and all variations within the equivalent scope should be interpreted as being included in the invention. Explanation of the symbols

[0076] 10. Framework Attributes / Features Input Section 20. User Specification Input Section 30. Deep Learning Framework Recommendation Section 40. Recommendation Framework Information Output Section

Claims

Claim 1 A framework attribute / feature input unit for inputting framework attribute information including features, feature values ​​of said features, scores of said feature values, and weights of said features for each of a plurality of deep learning frameworks and a model template for framework selection; a user specification input unit for inputting a user specification including an input deep learning model to be converted and data attribute information to be used, and including at least one of a project objective, a target device or backend environment, training data, and a user algorithm written in pseudocode; a model data analyzer that analyzes whether there are syntax errors in said user specification and the validity of the input data, and extracts a user algorithm from said user specification using a procedure analyzer and a function analyzer, analyzes the relationship between said extracted user algorithm and the input data to select a plurality of features, determines whether each of said a plurality of deep learning frameworks supports the feature values ​​of said selected features, calculates a feature-specific score by multiplying the score of said feature value and the weight of said feature for the supported features, and calculates a framework score by summing the feature-specific scores, adds a deep learning framework that supports all said selected multiple features to a list of supported frameworks, and the highest framework among said list of supported frameworks A deep learning framework recommendation system based on feature selection, characterized by comprising: a deep learning framework recommendation unit including a framework selector that recommends a deep learning framework having a score; and a recommendation framework information output unit that outputs information about the recommended deep learning framework and a framework optimization model design. Claim 2 A feature selection-based deep learning framework recommendation system according to claim 1, wherein the user specification is composed of a text file in a user-defined format and further includes at least one of the purpose of the program, a device to perform deep learning, a model size, and a data type. Claim 3 A deep learning framework recommendation method based on feature selection, performed by a deep learning framework recommendation system based on feature selection, comprising: a step in which a user specification input unit receives a user specification; a step in which a model data analyzer analyzes whether there are syntax errors in the user specification and the validity of the input data; a step in which a framework selector extracts a user algorithm from the user specification using a procedure analyzer and a function analyzer; a step in which the framework selector analyzes the relationship between the extracted user algorithm and the input data to select a plurality of features; a step in which the framework selector determines whether each of the plurality of deep learning frameworks supports the feature value of each selected feature, calculates a feature-specific score by multiplying the score of the feature value and the weight of the feature for the supported features, and then calculates a framework score by summing the feature-specific scores; and a step in which the framework selector adds a deep learning framework that supports all of the selected plurality of features to a list of supported frameworks and recommends a deep learning framework having the highest framework score among the list of supported frameworks. A deep learning framework recommendation method based on feature selection, characterized by including the step of a recommendation framework information output unit outputting information about the recommended deep learning framework and a framework optimization model design.

Images