CN115563060A - Visual management system for graph neural network - Google Patents
Visual management system for graph neural network Download PDFInfo
- Publication number
- CN115563060A CN115563060A CN202211151054.9A CN202211151054A CN115563060A CN 115563060 A CN115563060 A CN 115563060A CN 202211151054 A CN202211151054 A CN 202211151054A CN 115563060 A CN115563060 A CN 115563060A
- Authority
- CN
- China
- Prior art keywords
- graph
- model
- neural network
- module
- log
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/10—File systems; File servers
- G06F16/16—File or folder operations, e.g. details of user interfaces specifically adapted to file systems
- G06F16/168—Details of user interfaces specifically adapted to file systems, e.g. browsing and visualisation, 2d or 3d GUIs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/10—File systems; File servers
- G06F16/11—File system administration, e.g. details of archiving or snapshots
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/10—File systems; File servers
- G06F16/18—File system types
- G06F16/1805—Append-only file systems, e.g. using logs or journals to store data
- G06F16/1815—Journaling file systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a visual management system for a graph neural network, which comprises a graph data set analysis and visual module, a model structure building and training module, a log management module and a parameter tuning module. And the storage standard of the graph data set is normalized, the graph data set is visually displayed by using a visual library, and the graph data set is contrasted. The neural network model of the modular construction diagram realizes flexible calling, self-defining, training and sharing, and simplifies the development steps. Recording the operation of the graph neural network model, recording the trained model parameters and establishing a log, carrying out normalized storage through the log, carrying out visual comparative analysis, carrying out parameter optimization on the model on the basis, recording the result, and calling the optimized model parameters at any time. An application programming interface is constructed to be connected with each module, the display graph data set and the optimized graph neural network model are called at any time according to user requirements, a training task or model evaluation is implemented, and development efficiency is improved.
Description
Technical Field
The invention relates to the technical field of a graph neural network, in particular to a visual management system for the graph neural network.
Background
There is a lot of graph structure data in the current stage of research work, such as: social networks, traffic networks, criminal group organizations, protein networks and the like, the graph data are closer to the network topological structure relationship in the real scene, and richer information is contained. In recent years, a Graph Neural Network (GNN) model is rapidly developed, which can comprehensively mine topological structure information of graph data and node attribute information, and has many successful application examples in multiple fields such as finance anti-fraud, social networks, user recommendation, biomedicine and the like.
The training process of the graph neural network model generally comprises loading a graph structure data set, constructing the graph neural network model, training a graph neural network model log, and analyzing and adjusting parameters of the graph neural network so as to train an optimal graph neural network model for a certain data set. The prior art lacks a scheme capable of performing persistent storage and management, and therefore, a management system capable of effectively storing graph data and managing a graph neural network model is urgently needed, and training and evaluation processes are simplified.
Disclosure of Invention
The embodiment of the invention provides a visual management system for a graph neural network, which is used for eliminating or improving one or more defects in the prior art and solving the problem that a graph data set cannot be efficiently managed and the graph neural network cannot be called for training and analyzing in the prior art.
The technical scheme of the invention is as follows:
in one aspect, the present invention provides a visualization management system for a graph-oriented neural network, including:
the graph data set analysis and visualization module is used for acquiring one or more graph data sets uploaded by a user, storing the graph data sets according to a set storage standard, generating a graph data set information table and visually displaying each graph data set based on a preset visualization library;
the model structure building and training module comprises an in-layer module, an interlayer module and a model parameter module, and is used for building one or more graph neural network models in a blocking manner, calling a specified graph data set to train the specified graph neural network models and generating a model information table;
the log management module is used for recording the logs of uploading, modifying, deleting, training and calling each graph neural network model, recording parameter values of the graph neural network model after corresponding operation by each log, and generating a log information table according to each log for management and visual analysis;
the parameter tuning module is used for performing parameter tuning on the neural network models of each graph and generating a tuning information table;
the graph data set analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module are connected through an application programming interface.
In some embodiments, the graph data set analysis and visualization module divides each graph data set into a training set, a test set, and a validation set in a set proportion.
In some embodiments, the graph dataset analysis and visualization module is further to: and extracting graph structure characteristic information of each graph data set by adopting NetworkX and DGL and storing the graph structure characteristic information in a MongoDB database.
In some embodiments, the graph dataset analysis and visualization module is further to: and (3) carrying out visual display on the graph structure characteristic information of each graph data set by adopting Echarts through a table, a node degree distribution line graph and/or a graph part node connection force graph.
In some embodiments, the graph dataset analysis and visualization module is further to: and acquiring the graph structure characteristic information corresponding table, the node degree distribution line graph and/or the difference part of the graph part node connection force graph of each graph data set, and highlighting.
In some embodiments, the model structure building and training module is used for performing modular management on the neural network models of the figures, and at least records the model name of each neural network of the figures, a figure data set related to each neural network of the figures, module information contained in each neural network of the figures, parameters of each neural network of the figures and training result indexes of each neural network of the figures;
the module information comprises one or more pieces of sub-module information forming each graph neural network, an association structure among the sub-modules, and default values and parameter ranges of the sub-modules.
In some embodiments, the model structure building and training module is further configured to:
receiving a parameter setting request of a user for a first designated graph neural network model, and storing and calling preset parameters of the user through a Json shared file so as to train the first designated graph neural network model on the basis of the preset parameters.
In some embodiments, the log management module is further configured to:
acquiring one or more parameter items to be analyzed selected by a user aiming at a second designated graph neural network model;
searching the log information table and the tuning information table, searching training batches with different to-be-analyzed parameter items and the same other parameter items in the second specified graph neural network model based on a control variable mode, obtaining model training effects of the corresponding training batches, and comparing and evaluating influence relations of the to-be-analyzed parameter items on the training effects of the second specified graph neural network model.
In some embodiments, the log management module is further configured to:
inquiring the log information table to split key value pair information of each parameter default parameter value in each log in a database; if the log has model parameter loss, calling default parameter values according to corresponding key values to fill in the information, and updating the log;
and optimizing and visualizing the log information table according to a set rule.
In some embodiments, the parameter tuning module is further configured to: and generating a tuning target curve according to the tuning information table for visual presentation.
The invention has the beneficial effects that:
the visual management system for the graph neural network standardizes the storage standard of the graph data set through the graph data set analysis and visualization module, visually displays the graph data set by using the visualization library, and compares the differences of the graph data set. The method comprises the steps of building a model structure and building a graph neural network model in a modularization mode through a training module to achieve flexible calling, self-defining, training and sharing, simplifying development steps, recording operation on the graph neural network model through a log management module, recording model parameters after training and building a log, performing standardized storage through the log, performing visual comparative analysis, performing parameter optimization on the model on the basis, recording results, and calling optimized model parameters at any time. An application programming interface is constructed to be connected with each module, the display graph data set and the optimized graph neural network model are called at any time according to user requirements, a training task or model evaluation is implemented, and development efficiency is improved.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
It will be appreciated by those skilled in the art that the objects and advantages that can be achieved with the present invention are not limited to the specific details set forth above, and that these and other objects that can be achieved with the present invention will be more clearly understood from the detailed description that follows.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:
fig. 1 is a schematic structural diagram of a visualization management system facing a graph neural network according to an embodiment of the present invention.
Fig. 2 is a system architecture diagram of a visualization management system facing a neural network according to an embodiment of the present invention.
Fig. 3 is a system technical flowchart of a visualization management system facing a graph neural network according to an embodiment of the present invention.
Fig. 4 is a flowchart illustrating a log uploading process in a visualization management system of a graph neural network according to an embodiment of the present invention.
Fig. 5 is a flowchart illustrating a use of a log management module in a visualization management system of a neural network according to an embodiment of the present invention.
Fig. 6 is a parameter tuning flow chart of a parameter tuning module in a visualization management system of a graph neural network according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings. The exemplary embodiments and descriptions of the present invention are provided to explain the present invention, but not to limit the present invention.
It should be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and/or processing steps closely related to the scheme according to the present invention are shown in the drawings, and other details not so relevant to the present invention are omitted.
It should be emphasized that the term "comprises/comprising" when used herein, is taken to specify the presence of stated features, elements, steps or components, but does not preclude the presence or addition of one or more other features, elements, steps or components.
It is also noted that, unless otherwise specified, the term "coupled" is used herein to refer not only to a direct connection, but also to an indirect connection with an intermediate.
At present, the graph structure data set management system cannot support a user to upload a custom data set and does not support the comparison between two data sets. The construction and training of the graph neural network model are generally written and constructed in special code writing software. The operations of graph data set analysis, model building, log analysis, automatic parameter tuning and the like are processed in a separated mode, and a user needs to go to other platforms to analyze or write codes to realize functions, so that the operation and training processes of the user on the graph neural network model are not coherent, and the cooperation of all modules cannot be used for helping the user to research better.
In addition, the existing graph neural network training platform lacks log management on user model training, log storage and log management services cannot be provided, and a user can only manually store log files and write scripts to analyze relevant information. The existing deep learning log visualization system can only visualize log information after the log information is deployed, can only visualize logs of a single model, does not distinguish a graph neural network model structure from an adjusting and optimizing parameter module and the like, and does not support difference comparison among multi-model logs.
After the user analyzes the log of the graph neural network model, the user may choose to manually adjust the parameters to retrain the model or choose specific parameters to perform automatic parameter tuning of the model, the existing automatic parameter tuning function can only perform parameter tuning of a black box, if the middle is interrupted, the model can only restart, the tuning time of the neural network model is long, and the automatic parameter tuning of the graph neural network model is time-consuming and poor in stability.
The disadvantages of the prior art can be summarized as:
1) Graph structure features that do not support automated visualization of user-defined datasets do not allow visualization of differences between structures of contrasting graph datasets in log visualization analysis.
2) In the prior art, only the parameter result visualization of the relevant logs is supported after the log files are uploaded by a user, the model log visualization with continuity is not supported, the differences among models, data sets, structural parameters, training parameters and the like cannot be structurally distinguished, and the screening of the combination differences of a plurality of parameters is not supported, for example: a plurality of parameters of different models are used as a group to compare the influence of different combinations on model training, and the existing system does not support the function of visualizing the model log grouping.
3) In the existing system, a user cannot quickly perform automatic parameter tuning of a related model after analyzing the influence of log parameters, and meanwhile, the automatic parameter tuning of the existing deep neural network model cannot be visually monitored and the tuning speed is low.
4) The log management and the model training of the existing management system are separated, and the log management and the model training cannot be interacted with each other, so that the efficiency of a process of forming optimal parameters by a user in research and development of model combination is low, and the consistency is poor.
Therefore, the present invention provides a visualization management system for a graph neural network, as shown in fig. 1, including: the system comprises a graph data set analysis and visualization module, a model structure building and training module, a log management module and a parameter tuning module, wherein the graph data set analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module are connected through an application programming interface.
The graph data set analysis and visualization module is used for acquiring one or more graph data sets uploaded by a user, storing the graph data sets according to a set storage standard, generating a graph data set information table and visually displaying each graph data set based on a preset visualization library.
The model structure building and training module comprises an in-layer module, an interlayer module and a model parameter module, and is used for building one or more graph neural network models in a blocking mode, calling a specified graph data set to train the specified graph neural network models and generating a model information table.
The log management module is used for recording the logs of uploading, modifying, deleting, training and calling of the neural network models of the various figures, recording parameter values of the neural network models of the figures after corresponding operations in the various logs, and generating a log information table according to the various logs for management and visual analysis.
And the parameter tuning module is used for performing parameter tuning on the neural network models of the graphs and generating a tuning information table.
In this embodiment, as shown in fig. 2, the system architecture formed by the modules can be expressed as 5 layers, which are: the device comprises a hardware layer, a persistence layer, a model layer, a back-end layer and a front-end layer.
In the hardware layer, the graph data set analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module can adopt electronic equipment which can store and execute computer programs, such as a computer processor or a single chip microcomputer, and can adopt a Central Processing Unit (CPU) in the practical application process, and the model structure building and training module can adopt a Graphic Processing Unit (GPU) due to the fact that the training of the graph neural network model is involved.
In this embodiment, the graph data set information table generated by the graph data set analysis and visualization module is used for storing the structural feature information of the graph data set; storing a model information table generated by the model structure building and training module, wherein the model information table is used for storing the structure definition information of the model; the log information table of the log management module is used for storing training log information related to the model; and the tuning information table generated by the parameter tuning module is stored and used for storing the information of the automatic tuning process.
The model layer can be realized based on an existing Graph neural network development technology Library, such as a DGL (Deep Graph Library) Library, a NetworkX or a Pyroch, and comprises a Graph structure information extraction module (namely a Graph data set analysis and visualization module), a Graph neural network model training module (namely a model structure building and training module) and an automatic parameter tuning tool (namely a parameter tuning module). In the graph structure information extraction module, after a user uploads a data set, a graph is constructed through DGL and NetworkX, and relevant features of a graph structure are automatically extracted and stored in a graph data information table.
The graph neural network model training module realizes the conventional graph convolution neural network (GCN), the graph attention neural network (GAT), the GraphSage and the like, and in addition, the GNN design space is modularized and componentized and divided into: the system comprises an Intra-layer Module, an Inter-layer Module and a model parameter Module, wherein a user can design and combine different sub-modules, set a parameter space and generate and train a new graph neural network model.
The automatic parameter tuning tool is based on the automatic parameter tuning tool kit as follows: the graph neural network model realized by the Hyperopt parameter tuning package combined system starts automatic parameter tuning of the graph neural network model and stores tuning state information into a tuning information table by using a Netstat real-time monitoring tuning process.
At the back end layer, the open framework based on python configures the route responding to the front end through the resource locator, such as Django framework: and (5) returning a response view to the route definition related function of the resource locator by using a view function (views), and designing a front-end HTML webpage template by using the template.
The code function can adopt Pymnogo in Python to flexibly add, delete, change and check data in the MongoDB database. The back-end layer can realize the message transmission of related parameters in the form of Json shared files, and start a new process training graph neural network model by Popen in the Subpprocesses and monitor the returned process numbers. And the front end and the back end carry out data interaction through Ajax.
The front-end layer can use HTML, bootStrap, javaScript, JQuery, etc. to complete the design of the interface. The Jupiter notewood and the IFrame are used for embedding the system interface into the code writing of the Jupiter notewood, so that a user can use the system flexibly and conveniently.
The front-end layer mainly designs seven interfaces:
the graph data set uploading/structure visualization interface is used for being matched with a graph data set analysis and visualization module to realize the increase, deletion, modification and examination of the graph data set and the visualization of the extracted structure characteristics of the graph data set.
And secondly, the graph data set comparison interface is used for adapting to a graph data set analysis and visualization module to realize the comparison of the difference of the structural features of different graph data sets and the calculation of the similarity.
And thirdly, the model information management interface is used for adapting to the model structure building and training module to realize the definition and structure management of the graph neural network model structure information.
And fourthly, the model training interface is used for adapting to the model structure building and training module to realize the rapid combination and model training of the graph neural network model.
And fifthly, a log information management analysis interface is used for managing a log management module, training log information of a persistent and standardized storage model, and supporting fast log parameter screening inquiry, log grouping and log comparison.
And sixthly, an automatic tuning monitoring interface is used for managing the parameter tuning module, supporting starting of automatic parameter tuning and monitoring tuning in real time.
And seventhly, the Jupyter notewood interface is used for self-defining and constructing an Application Programming Interface (API) interface, the IFrame is used for embedding the system into code writing, a graph data set training interface, a model structure definition interface, a model training and log management interface and an automatic tuning interface are opened, and the conventional process corresponding to the user for researching the graph neural network model is realized.
The following describes the visualization management system for the neural network of the graph according to this embodiment based on a development process of the neural network model, and the specific process includes: graph data set analysis and visualization → graph neural network model structure definition training → model training log management and visualization analysis → model automatic parameter tuning and monitoring. The system constructed by the method embeds the related API into Jupitter notewood through IFrame, and directly calls the corresponding system interface API in code writing.
The system comprises the following execution steps:
the first step is as follows: the graph data set graph structure characteristics are visualized, so that a user can visualize the graph data structure characteristics before developing a model, analyze and compare differences of different graph data structure characteristics, and simultaneously can visually analyze and compare influences of different data set graph structure characteristics on logs during log visualization analysis.
The second step is that: the definition of the graph neural network model structure not only defines the graph neural network model for training, but also can define a log management structure aiming at the structural characteristics of the corresponding graph neural network model.
The third step: and performing log visual management, namely constructing a structured management log on the basis of the second step of model structure definition, and simultaneously supporting the comparison of parameters of different graph neural network models and the functional comparison of parameter grouping.
The fourth step: and the automatic parameter tuning is carried out, so that a user can rapidly carry out automatic parameter tuning after analyzing the parameters of the neural network model of the graph, visual automatic parameter tuning monitoring is carried out on the basis of a log management function, and the automatic parameter tuning process is accelerated by a pre-stored and managed historical tuning information table.
In the four steps, a Jupitter Notebook is embedded in the system to construct a quick interface of the function, so that a user can conveniently and quickly call related functions in the development process of the graph neural network model code, the interactivity with the user is improved, and the development and research efficiency is improved.
Specifically, as shown in fig. 3, the graph data set analysis and visualization module implements graph data set analysis and visualization functions, the front layer of the system supports a user to upload a graph data set, the back layer persistently and normalizes the data set uploaded by the user and stores the data set in a designated file, and normalizes and manages a path, and a training set, a test set and a verification set are automatically partitioned. Specifically, the graph data set analysis and visualization module divides each graph data set into a training set, a test set and a verification set according to a set proportion.
The Graph data set analysis and visualization module constructs a Graph from a Graph data set based on NetworkX and DGL (Deep Graph Library is a framework for Deep learning on the Graph), extracts Graph structure characteristic information of each Graph data set and stores the Graph structure characteristic information into a Graph data information table in the MongoDB.
The graph data set analysis and visualization module also adopts Echarts to visually display the graph structure characteristic information of each graph data set through a table, a node degree distribution line graph and/or a graph part node connection force line graph. The differences between the two image datasets are automatically compared and the differences between the dataset features are displayed at the front end by highlighting.
As shown in fig. 3, the model structure building and training module is used for realizing model structure building and training, and a rear-end layer of the system can build various conventional graph neural network models GCN, GAT, graphSage and the like based on DGL. A general analysis graph neural network model (GNN) model structure space can be modularized and componentized, and comprises the following steps: the three main modules of the intra-layer module, the inter-layer module and the model parameter module support a user to combine the modules to generate a plurality of variants of the GNN model, so that the GNN model suitable for the own graph data set is found. In some embodiments, the model structure building and training module is used for performing modular management on the neural network models of the figures, and at least records the model name of each neural network of the figures, a figure data set related to each neural network of the figures, module information contained in each neural network of the figures, parameters of each neural network of the figures and training result indexes of each neural network of the figures; the module information comprises one or more pieces of sub-module information forming each graph neural network, an association structure among the sub-modules, and default values and parameter ranges of the sub-modules. Model structure is built and is trained the module, still is used for: and receiving a parameter setting request of a user for the first designated graph neural network model, and storing and calling preset parameters of the user through a Json shared file so as to train the first designated graph neural network model on the basis of the preset parameters.
The model structure building and training module builds a model information table to manage the characteristic information related to the model, and the table divides the model into: model name (ModelName), model-related graph dataset (DataName), model submodule information (Modules, which records submodule information constituting a graph neural network model, where each submodule includes a default value and a parameter value range), parameter information (Parameters, which records model-related Parameters, each parameter includes a default value and a value range), a training Result indicator (Result) output by the model, and possibly some other information (other) of the model.
The model structure building and training module manages any model in a structuralized and standardized mode through a model information table. When a user realizes a new model or adds a new module component in an existing model, the model definition and the module definition of the modification interface uploading model can be provided in a front-end system, and the system can automatically update the model module to a model information table for management and synchronize the model module to the subsequent log management function and the automatic tuning function.
In the process of training the GNN model, the system gives the GNN modularized interfaces at the front end, a user can directly set parameters at the front end to directly start self-defined GNN model training, the system can transmit the parameters input from the front end through the Json shared file, training of the corresponding model is started according to the received parameters through the supprocesses, and logs generated by model training in the system can be automatically stored in a log information table.
As shown in fig. 3, the log management module implements log management, and implements persistent management and visual analysis on the model training log in the log management module, and mainly includes four functions of log uploading, multi-model log keyword screening visualization, log grouping, and log comparison. The system manages and analyzes the model based on the cooperation of the model information table and the log information table, the log information table stores the log of the multi-graph neural network model, and each piece of data records one piece of training log data.
In some embodiments, the log management module is further to: inquiring a log information table to split key value pair information of each parameter default parameter value in each log in a database; if the log has model parameter loss, calling default parameter values according to corresponding key values to fill in the information, and updating the log; and optimizing and visualizing the log information table according to a set rule. Specifically, in the log management module, as shown in fig. 4, a log uploading function includes first calling an Init method to connect to a database MongoDB, calling an UpLoadLog method to upload string-type log information after connecting to the database, automatically splitting key-value pair information of log parameters by the system, querying parameters defined by a corresponding model in a model information table according to a ModelName in each log, if some parameter information is missing in the uploaded log, automatically filling corresponding default parameter values in a default model information table, if the parameter information is not missing, filling corresponding parameter values of the extracted log, and finally storing complete model log information in the log information table. Thus, each log is in one-to-one correspondence with the parameters of the corresponding model in the model information table, so that the logs are managed in a standardized way. The system simultaneously develops a tool packet for log uploading, and when a user trains a self-defined model in a local compiling environment, the tool packet can be called in the code, and the log trained by the model is uploaded to a log information table in real time.
The system realizes the visualization of the log information, as shown in fig. 5, a model list which can be selected to view is given through a model information table, and a user can select a model to view at a front-end layer. If the user selects a single model, the system inquires and returns the corresponding model parameter information in the model information table; if a plurality of models are selected, the system takes out the common Parameters of the corresponding models by searching the Parameters of the corresponding models in the model information table and performing intersection, and returns the common Parameters, wherein the common Parameters can comprise five groups of Parameters, namely DataName, modules, parameters, results and other.
After the relevant parameters of the single or multiple models selected by the user are returned, the user selects the log parameters to be checked from the parameters, the system inquires the corresponding training logs in the log information table according to the parameters selected by the user and returns the log information table of the corresponding inquiry parameters for visualization, so that the user can accurately and conveniently pay attention to the contents of the log parameters researched by the user, and redundant information is removed. After the front end displays the summary parameter information of the user check model, the user can click a certain log to directly enter the detailed information for checking the log.
In some embodiments, the log management module is further to: and acquiring one or more parameter items to be analyzed selected by the user aiming at the second designated graph neural network model. And searching training batches with different parameter items to be analyzed and the other parameter items consistent in the second specified graph neural network model based on a control variable mode by searching the log information table and the tuning information table, and obtaining model training effects of the corresponding training batches so as to compare and evaluate the influence relationship of the parameter items to be analyzed on the training effects of the second specified graph neural network model.
Specifically, the system supports a user to select parameters to group logs, after the user selects parameters to be analyzed, the system groups the logs according to the model parameters recorded in the model information table and the trained model parameters recorded in the log information table, divides the logs with the same values of other parameters except the parameters to be analyzed into the same group, and distinguishes different groups through different colors at the front end layer, so that the effect on a model result when one or more parameters to be analyzed change is quickly checked. The user selects a plurality of logs for comparison, the system jumps to a log comparison interface, the system analyzes and compares the difference of the selected plurality of logs, gives the difference of specific parameters in the logs and the specific information of each log data, highlights the parameter value of the difference, and the user can click to check the difference between the log result and the model parameter of the optimal log result.
As shown in fig. 3, the parameter tuning module is used to automatically tune parameters of the neural network model, and the automatic parameter tuning function of the neural network model is designed and developed on the basis of the Hyperopt toolkit, as shown in fig. 6, parameters of a certain model in the model information table are divided into two types: tuning parameters and default parameters. The tuning parameters are parameters needing tuning by a tuning algorithm, and a user can define parameter tuning ranges; the default parameter is a parameter which does not need to be adjusted and optimized, and the parameter value supports the user to define a parameter value, and the default parameter value in the model information table can also be directly used.
When the user starts the parameter tuning function, a tuning log is generated for the tuning process and stored in the table, and the system records the tuning process and monitors whether the process is running or not. During tuning, the system can select a value from the tuning range of each tuning parameter to combine with a default parameter to generate a model parameter combination and start the training of a model, before starting the training of the model, the system can check whether a trained model log exists in a log information table, if so, the result information of the log is directly called to realize accelerated tuning, otherwise, the model of the parameter combination is started to train and store the log result.
The training logs of each model tuning in the automatic parameter tuning process are stored and visualized on a front-end interface, and a user can check detailed information of tuning target curves and related model parameter logs generated in each tuning training process at any time.
Further, in this embodiment, an API is developed by embedding Jupyter noteebook, a Jupyter noteebook compiling interface is embedded in the system interface, and three interface interfaces are implemented by IFrame: a data set interface, a model structure definition interface, and a log management and automatic parameter tuning interface. When a user compiles and realizes a custom model, an interface can be called when the Jupiter notewood code is compiled, wherein the data set interface is used for uploading and viewing the data structure characteristics of a graph by a visual data set; the model structure defining interface is used for visually defining a network model structure and starting model training; the model training and log management is used for visually analyzing historical model training logs in model training; the automatic parameter tuning interface is used for realizing an optimal parameter combination model by automatic parameter tuning. Finally, the management and analysis of the graph neural network model and the log visualization while writing codes are realized.
In some embodiments, the parameter tuning module is further configured to: and generating a tuning target curve according to the tuning information table for visual presentation.
The invention constructs an integrated and interfaced system platform with visualization of a graph structure data structure, combined training of a graph neural model module, real-time multi-model and multi-log management and efficient analysis and comparison. Compared with the traditional method for developing the graph neural network model, the method has the advantages that the model development and training are carried out singly, the graph data visualization, the model training log management and the model tuning are combined into the graph neural network model development, a platform is provided for a user, a convenient interface is realized, the function of each module is called quickly in the user development process, and the development efficiency is improved.
The method realizes the structured management of the graph neural network model by multi-model training log analysis oriented to the graph neural network model, and expresses the model by a structured module of system design. And the management of standardized persistence of the model training log is realized on the basis of model structural management. The method supports the management of logs of various models, quickly compares the similarities and differences among the training log results of specific modules of various models, realizes visual analysis and comparison, and provides convenience for the analysis and development of user models.
The model structured management and the log structured management are combined, automatic tuning monitoring and visual monitoring are achieved, and the tuning process is accelerated by inquiring historical logs. The method overcomes the defects that the automatic parameter tuning process of the conventional deep neural network model cannot be monitored and is slow in speed.
In summary, the visualization management system for the graph neural network of the present invention standardizes the storage standard of the graph data set through the graph data set analysis and visualization module, and visually displays the graph data set by using the visualization library to compare the graph data set differences. The method comprises the steps of building a model structure and building a graph neural network model in a modularization mode through a training module to achieve flexible calling, self-defining, training and sharing, simplifying development steps, recording operation on the graph neural network model through a log management module, recording model parameters after training and building a log, performing standardized storage through the log, performing visual comparative analysis, performing parameter optimization on the model on the basis, recording results, and calling optimized model parameters at any time. An application programming interface is constructed to be connected with each module, the display graph data set and the optimized graph neural network model are called at any time according to user requirements, a training task or model evaluation is implemented, and development efficiency is improved.
Those of ordinary skill in the art will appreciate that the various illustrative components, systems, and methods described in connection with the embodiments disclosed herein may be implemented as hardware, software, or combinations thereof. Whether this is done in hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an Application Specific Integrated Circuit (ASIC), suitable firmware, plug-in, function card, or the like. When implemented in software, the elements of the invention are the programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine-readable medium or transmitted by a data signal carried in a carrier wave over a transmission medium or a communication link. A "machine-readable medium" may include any medium that can store or transfer information. Examples of a machine-readable medium include an electronic circuit, a semiconductor memory device, a ROM, a flash memory, an Erasable ROM (EROM), a floppy disk, a CD-ROM, an optical disk, a hard disk, an optical fiber medium, a Radio Frequency (RF) link, and so forth. The code segments may be downloaded via computer networks such as the internet, intranet, etc.
It should also be noted that the exemplary embodiments mentioned in this patent describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, may be performed in an order different from the order in the embodiments, or may be performed simultaneously.
Features that are described and/or illustrated with respect to one embodiment may be used in the same way or in a similar way in one or more other embodiments and/or in combination with or instead of the features of the other embodiments in the present invention.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention, and various modifications and changes may be made to the embodiment of the present invention by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (10)
1. A visualization management system for a graph-oriented neural network, comprising:
the graph data set analysis and visualization module is used for acquiring one or more graph data sets uploaded by a user, storing the graph data sets according to a set storage standard, generating a graph data set information table and visually displaying each graph data set based on a preset visualization library;
the model structure building and training module comprises an in-layer module, an interlayer module and a model parameter module, and is used for building one or more graph neural network models in a blocking manner, calling a specified graph data set to train the specified graph neural network models and generating a model information table;
the log management module is used for recording the logs of uploading, modifying, deleting, training and calling each graph neural network model, recording parameter values of the graph neural network model after corresponding operation by each log, and generating a log information table according to each log for management and visual analysis;
the parameter tuning module is used for carrying out parameter tuning on the neural network models of the graphs and generating a tuning information table;
the graph data set analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module are connected through an application programming interface.
2. The visualization management system of claim 1, wherein the graph data set analysis and visualization module divides each graph data set into a training set, a testing set, and a verification set according to a set proportion.
3. The neural network-oriented visualization management system of claim 1, wherein the graph dataset analysis and visualization module is further configured to:
and extracting graph structure characteristic information of each graph data set by adopting NetworkX and DGL and storing the graph structure characteristic information in a MongoDB database.
4. The neural network-oriented visualization management system of claim 3, wherein the graph dataset analysis and visualization module is further configured to:
and (3) adopting Echarts to visually display the graph structure characteristic information of each graph data set through a table, a node degree distribution line graph and/or a graph part node connection force graph.
5. The neural network-oriented visualization management system of claim 4, wherein the graph dataset analysis and visualization module is further configured to:
obtaining a graph structure characteristic information corresponding table of each graph data set,
The node degree distribution line graph and/or graph part nodes are connected with the difference part of the force graph and highlighted.
6. The visualization management system for the neural networks of the figures according to claim 1, wherein the model structure building and training module is configured to perform modular management on the neural network models of the figures, and record at least a model name of each neural network of the figures, a figure data set related to the neural networks of the figures, module information included in the neural networks of the figures, parameters of the neural networks of the figures, and a training result index of the neural networks of the figures;
the module information comprises one or more pieces of sub-module information forming each graph neural network, an association structure among the sub-modules, and default values and parameter ranges of the sub-modules.
7. The visualization management system of claim 1, wherein the model structure building and training module is further configured to:
receiving a parameter setting request of a user for a first designated graph neural network model, and storing and calling preset parameters of the user through a Json shared file so as to train the first designated graph neural network model on the basis of the preset parameters.
8. The neural network-oriented visualization management system of claim 1, wherein the log management module is further configured to:
acquiring one or more parameter items to be analyzed selected by a user aiming at a second designated graph neural network model;
searching the log information table and the tuning information table, searching training batches with different to-be-analyzed parameter items and the same other parameter items in the second specified graph neural network model based on a control variable mode, obtaining model training effects of the corresponding training batches, and comparing and evaluating influence relations of the to-be-analyzed parameter items on the training effects of the second specified graph neural network model.
9. The neural network-oriented visualization management system of claim 1, wherein the log management module is further configured to:
inquiring the log information table to split key value pair information of each parameter default parameter value in each log in a database; if the log has model parameter loss, calling default parameter values according to corresponding key values to fill in the information, and updating the log;
and optimizing and visualizing the log information table according to a set rule.
10. The visualization management system of claim 1, wherein the parameter tuning module is further configured to:
and generating a tuning target curve according to the tuning information table for visual presentation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211151054.9A CN115563060A (en) | 2022-09-21 | 2022-09-21 | Visual management system for graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211151054.9A CN115563060A (en) | 2022-09-21 | 2022-09-21 | Visual management system for graph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115563060A true CN115563060A (en) | 2023-01-03 |
Family
ID=84740791
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211151054.9A Pending CN115563060A (en) | 2022-09-21 | 2022-09-21 | Visual management system for graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115563060A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116977525A (en) * | 2023-07-31 | 2023-10-31 | 之江实验室 | Image rendering method and device, storage medium and electronic equipment |
| CN117785177A (en) * | 2023-12-26 | 2024-03-29 | 北京汉勃科技有限公司 | Visual large model fine tuning method and device |
-
2022
- 2022-09-21 CN CN202211151054.9A patent/CN115563060A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116977525A (en) * | 2023-07-31 | 2023-10-31 | 之江实验室 | Image rendering method and device, storage medium and electronic equipment |
| CN116977525B (en) * | 2023-07-31 | 2024-03-01 | 之江实验室 | Image rendering method and device, storage medium and electronic equipment |
| CN117785177A (en) * | 2023-12-26 | 2024-03-29 | 北京汉勃科技有限公司 | Visual large model fine tuning method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115563060A (en) | Visual management system for graph neural network | |
| US10360504B2 (en) | Generalized faceted browser decision support tool | |
| US8185563B2 (en) | Data-visualization system and method | |
| US11593392B2 (en) | Transformation rule generation and validation | |
| EP2778929B1 (en) | Test script generation system | |
| KR102103902B1 (en) | Component-based machine learning automation device and method | |
| US9652497B2 (en) | Processing queries using hybrid access paths | |
| CN114816374B (en) | Visual data analysis flow modeling method and system | |
| US11556838B2 (en) | Efficient data relationship mining using machine learning | |
| US11681511B2 (en) | Systems and methods for building and deploying machine learning applications | |
| US20210012219A1 (en) | Dynamic generation of rule and logic statements | |
| US20210073655A1 (en) | Rule mining for rule and logic statement development | |
| CN117632710A (en) | Method, device, equipment and storage medium for generating test code | |
| US20210065016A1 (en) | Automatic generation of computing artifacts for data analysis | |
| Zerbato et al. | Supporting provenance and data awareness in exploratory process mining | |
| US11907238B2 (en) | Metadata-based software artifact search | |
| CN117829657A (en) | Spacecraft system evaluation system and evaluation method | |
| CN117009316A (en) | Method and system for managing full life cycle of data flow based on data blood margin | |
| CN116610558A (en) | Code detection method, device, electronic equipment and computer readable storage medium | |
| CN115983377A (en) | Automatic learning method, device, computing equipment and medium based on graph neural network | |
| CN115794858A (en) | Query statement processing method, device, equipment and storage medium | |
| CN115310011A (en) | Page display method and system and readable storage medium | |
| CN117555814B (en) | Software testing method and device, storage medium and terminal | |
| CN118070075B (en) | Data analysis method, device, electronic device and storage medium | |
| CN118296320B (en) | Equipment system capacity data analysis mining device and method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |