CN111752554A - Multi-model cooperation system and method based on model arrangement - Google Patents
Multi-model cooperation system and method based on model arrangement Download PDFInfo
- Publication number
- CN111752554A CN111752554A CN202010419296.6A CN202010419296A CN111752554A CN 111752554 A CN111752554 A CN 111752554A CN 202010419296 A CN202010419296 A CN 202010419296A CN 111752554 A CN111752554 A CN 111752554A
- Authority
- CN
- China
- Prior art keywords
- model
- standardized
- client
- arrangement
- scheduling
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/30—Creation or generation of source code
- G06F8/35—Creation or generation of source code model driven
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Stored Programmes (AREA)
Abstract
The invention discloses a multi-model cooperation system and a method based on model arrangement, wherein the system comprises a first client, a model management background and a model library: the model library is used for storing a plurality of standardized models; the first client is used for constructing model arrangement tree nodes, setting standardized models in the model base on the model arrangement tree nodes to construct a model arrangement tree, and binding the standardized models on the model arrangement tree nodes to the graphic processor; the model management background generates a model scheduling relation table according to the model arrangement tree, the standardized models on the arrangement tree nodes and the graphics processor bound by the models; the system provided by the invention is provided with an independent and universal model scheduling engine, and can be applied to various scenes of calculation by using the model by using an intuitive arrangement mode of the model arrangement tree, and the system is not required to be realized by modifying codes when the model is updated or replaced.
Description
Technical Field
The present invention relates to the field of industrial intelligence, and more particularly, to a multi-model collaboration system and method based on model orchestration.
Background
At present, the development process of machine learning application basically comprises the steps of preparing data, selecting a model, adjusting an optimization model, and finally, enabling the model to be online for external application calling. Model application development is usually performed based on a shaft type method, technologies and code logics of several layers of scene-based data input and output processing, model scheduling logics, model algorithms, algorithm execution frameworks and the like are strongly bound to form an integral module which can not be divided, and therefore all calculation, models, scheduling and data of the whole application scene are bound together. However, such a model application development scheme has the following disadvantages:
(1) the data model based on the business scene modeling is mixed with the picture data and the algorithm, so that once the business modeling is changed, the scheduling logic of the algorithm and the algorithm need to be changed, errors are easy to occur, and the development efficiency is low;
(2) the whole scheme is bound with a specific model execution framework (mask or mmdecision), the upgrading and modification difficulty is high, the cost is high, and the technical background of rapid development and change of the existing deep learning model framework is difficult to adapt;
(3) the codes of the whole scheme are logically fused together, once a certain model or certain models need to be updated, all the models need to be shut down, and online updating and maintenance of the models cannot be realized;
(4) if a plurality of models participate in the computation together, the plurality of models are usually merged together for calling, or different models are separately called one by one through an API (application programming interface), when a part of models need to be replaced, only codes can be modified, and iteration of the whole version is performed;
(5) a general model scheduling engine is not formed independently, when a new model is taken, the model is directly integrated and used after being developed and trained, a plurality of places of codes are involved when the model needs to be updated or replaced, the efficiency is low, and meanwhile when the model is applied to a new scene every time, the whole scheme is redeveloped, tested and deployed from beginning to end, and manpower and time are consumed.
Disclosure of Invention
The invention aims to overcome at least one defect of the prior art and provides a multi-model cooperation system and method based on model arrangement, which are used for solving the problem that the existing models are inconvenient and fussy in the process of calling and replacing the models after being integrated and used.
The technical scheme provided by the invention is as follows:
a multi-model collaboration system based on model orchestration, the system comprising a first client, a model management backend, and a model repository:
the model library is used for storing a plurality of standardized models; the first client is used for constructing model arrangement tree nodes and arranging a plurality of standardized models stored in a model library on the model arrangement tree nodes to form a model arrangement tree; the first client is also used for binding a plurality of standardized models on all the model arrangement tree nodes to a graphics processor so that the plurality of standardized models run by the graphics processor;
the model management background generates a model scheduling relation table according to the model arrangement tree formed by the first client, the standardized model arranged on the node of the model arrangement tree and the graphics processor bound by the standardized model;
the client side uses the model arrangement tree to arrange a plurality of standardized models in the model base, wherein the standardized models refer to models with standardized input and output formats. The client adds model arranging tree nodes firstly, the nodes are sequentially connected in front and back or connected in parallel, the client can set standardized models on the model arranging tree nodes, the execution relation between the standardized models is arranged according to the relation between the model arranging tree nodes, and if the two nodes where the two standardized models are located are in a front-back sequential relation, the execution sequence of the two standardized models is sequentially executed in front and back. After the client lays out the standardized models by using the layout tree, the standardized models on the model layout tree need to be bound with the graphics processor for operating the standardized models. After the model arrangement tree is constructed at the client side by the model management background, a model scheduling relation table is generated according to the model arrangement tree, the standardized models on the nodes of the model arrangement tree and the graph processor bound by the model, the table comprises the tree structure of the model arrangement tree, the standardized models corresponding to each node in the model arrangement tree and the graph processor required when the model arrangement tree is executed.
The model scheduling relation table formed in the system provided by the invention records the arrangement information of the model tree, so that the model scheduling relation table can be used for subsequent model scheduling. When the client needs to update the existing model, only the standardized model needs to be updated in the model library, and when the client needs to change the scheduling sequence of the existing model, the arrangement sequence in the model arrangement tree can be directly changed, so that the system can be applied to various scenes for calculation by using the model, and the system does not need to be realized by modifying codes when the model is updated or replaced. Since the input and output of the model are standardized, the corresponding calling code does not need to be changed after the model is replaced. Moreover, the system enables the client to more intuitively control the execution sequence of the standardized models in a mode of arranging the model arrangement tree, and the arrangement result is more directly displayed through the arrangement tree.
Further, the system also comprises a second client and a model scheduling engine, wherein the second client is used for sending a visual detection task to the model scheduling engine;
and the model scheduling engine is used for acquiring and calling the standardized model from a model library to execute the visual detection task according to the model scheduling relation table generated by the model management background after receiving the visual detection task, and returning an execution result after the visual detection task to the second client.
The first client side is mainly used for arranging the standardized models to enable the model management background to generate the model scheduling relation table, and the second client side is mainly used for sending visual detection tasks to the model scheduling engine to call the standardized models after the model management background generates the model scheduling relation table. The system is provided with an independent model scheduling engine, after a visual detection task sent by a client is received, a model scheduling relation table of a model management background is obtained, a standardized model on a corresponding node is called according to a tree structure of a model arrangement tree in the table, namely a sequencing relation between each node, so that the visual detection task sent by a second client is completed, and after the task is completed by the model, the result is returned to the second client.
Further, the second client is configured to send a visual inspection task to the model scheduling engine, and specifically: the second client is used for sending the visual detection content and the model ID of the standardized model required to be called for executing the visual detection content to the model scheduling engine;
the model scheduling engine is configured to, after receiving the visual inspection content, obtain and call the standardized model from the model library according to the model scheduling relationship table generated by the model management background to execute the visual inspection task, and specifically includes: and the model scheduling engine is used for acquiring and calling a standardized model corresponding to the model ID from the model library to execute the visual detection content according to the model scheduling relation table generated by the model management background after receiving the visual detection content and the model ID.
The first client can set more than one standardized model on the model arrangement tree node, when sending the visual inspection task to the model scheduling engine, if the second client needs to call one of the standardized models on the model arrangement tree node, the standardized model can be called through the model ID, and the standardized model defines the parameter ID as the model ID when initializing the interface and is the unique identification of the standardized model.
Therefore, when the second client sends the visual detection task to the model scheduling engine, the visual detection task includes visual detection content and a model ID of a standardized model required for completing the visual detection content, and after the model scheduling engine acquires the tree structure of the model arrangement tree according to the background model scheduling relationship table, the model scheduling engine correspondingly calls a certain standardized model on the node of the model arrangement tree according to the model ID sent by the second client.
In the multi-model cooperative scheduling process, different models are frequently scheduled according to different inputs, a common processing mode is to use different models for arrangement, each model arrangement corresponds to one condition, and one standardized model on a node can be called in the actual calling process by arranging a plurality of standardized models on the nodes of the model arrangement tree, so that one model arrangement can meet the calling scenes of various models.
Further, the model base is a model mirror image base; the first client is also used for uploading a plurality of standardized models to the model management background; the model management background is further used for generating an image file of the standardized model after receiving the standardized model uploaded by the first client and storing the image file to the image warehouse;
the model scheduling engine is configured to obtain and call the standardized model corresponding to the model ID from the model library according to the model scheduling relationship table generated by the model management background to execute the visual inspection content, and specifically includes: and the model scheduling engine pulls and starts the image file of the standardized model corresponding to the model ID from the model image warehouse according to the model scheduling relation table generated by the model management background so as to execute the visual detection content.
The first client is also used for uploading and preparing the standardized model before arrangement, the standardized model can be uploaded to the model management background by the first client, after the standardized model is received by the model management background, the model uploaded by the first client is packaged into a mirror image file and pushed to a mirror image warehouse, the model is managed by the mirror image warehouse to share the model, meanwhile, the model is constructed into a container mirror image, the operation and maintenance difficulty of the model can be reduced, the deployment efficiency of the model at any end is improved, and the operation of the model is not influenced by the operation environment. Therefore, if the model scheduling engine needs to call the standardized model to complete the visual inspection task, the model scheduling engine needs to call the standardized model successfully by calling the model in the mirror image warehouse and pulling the mirror image file of the model to the local and starting the mirror image file.
Further, the first client is also used for uploading the running test task corresponding to the image file to a model management background;
the model management background is further used for starting the generated image file to execute the running test task and returning an execution result of the running test task to the first client after receiving the running test task corresponding to the image file and generating the image file of the standardized model.
The first client is further used for judging whether the image file successfully executes the running test task according to the execution result of the running test task after receiving the execution result of the running test task;
the model management background is further used for storing the image file to the model image warehouse when the first client determines that the image file successfully executes the operation test task.
After the first client uploads the standardized model, the standardized model is packaged into the image file by the model management background, and at the moment, the first client needs to test whether the uploaded model can normally run or not so as to ensure that the image file does not have problems in the subsequent calling process. Therefore, the first client needs to upload a test task to the model management background, the model background starts the image file of the standardized model to run the model after receiving the test task, the model outputs a result to the client after completing the test task, the first client judges whether the model successfully completes the test task, if so, the model management background pushes the image file to the image warehouse, and if not, the image file is failed to be pushed, and the first client can adopt measures such as re-uploading the model and the like to solve the problem.
Further, the model management background is also used for performing Restful packaging on the received interface of the standardized model after receiving the standardized model uploaded by the client and before generating the image file of the standardized model. The model management background is responsible for performing Restful packaging on the interface of the standardized model, so that the standardized model can be called remotely, and the service of the standardized model is realized.
Further, the first client and the second client are the same client. As can be seen from the above description of the technical solution, the first client is mainly used for uploading, testing and arranging the standardized model, that is, mainly used for preparing the standardized model before being called, and the second client is mainly used for uploading a model testing task to call the standardized model, that is, mainly used for calling the standardized model; when the second client considers that a new model or a new model arrangement needs to be replaced in the calling process, the first client can be used for re-uploading and arranging, and therefore when the two clients are the same client as an optimal scheme, a user can more conveniently and efficiently call and replace the model.
The technical scheme provided by the invention is as follows:
a multi-model cooperation method based on model arrangement is applied to a first client side and comprises the following steps: constructing model arrangement tree nodes, and arranging a plurality of standardized models in a model base on the model arrangement tree nodes to form a model arrangement tree; binding the standardized models provided on the model orchestration tree nodes to a graphics processor such that the standardized models are run with the bound graphics processor, the standardized models being stored in the model library.
The technical scheme provided by the invention is as follows:
a multi-model cooperation method based on model arrangement is applied to a second client side and comprises the following steps:
sending a visual inspection task to a model scheduling engine so that the model scheduling engine acquires and calls a standardized model from a model library according to a model scheduling relation table to execute the visual inspection task, wherein the model scheduling relation table is generated by a model management background according to a model arrangement tree formed by a first client, the standardized model set by the first client on a node of the model arrangement tree, and a graph processor bound to the standardized model by the first client; the standardized model is stored in the model library and is operated by the bound image processor;
and receiving the execution result of the visual detection task returned by the model scheduling engine.
The technical scheme provided by the invention is as follows:
a multi-model cooperation method based on model arrangement is applied to a model management background and comprises the following steps: and generating a model scheduling relation table according to a model arrangement tree constructed by a first client, a standardized model arranged on a node of the model arrangement tree by the first client and a graphic processor bound to the standardized model by the first client, so that a model scheduling engine calls the standardized model from a model library according to the model scheduling relation table to execute a visual detection task after receiving the visual detection task sent by a second client, and returns an execution result of the visual detection task to the second client.
The nodes of the model arrangement tree are constructed by the first client, the standardized models are arranged on the constructed nodes of the model arrangement tree by the first client, the model arrangement tree is formed by the nodes of the model arrangement tree in which the standardized models are arranged, the standardized models are operated by the bound image processor, and the standardized models are stored in the model library.
The technical scheme provided by the invention is as follows:
a multi-model cooperation method based on model arrangement is applied to a model scheduling engine and comprises the following steps: receiving a visual detection task sent by a second client, and acquiring a model scheduling relation table, wherein the model scheduling relation table is generated by a model arrangement tree constructed by a model management background according to a first client, a standardized model arranged on a node of the model arrangement tree by the first client, and a graphic processor bound to the standardized model by the first client, the standardized model is operated by adopting the bound image processor, and the standardized model is stored in a model library;
and acquiring and calling the standardized model in a model library according to the model scheduling relation table to execute the visual detection task, and returning an execution result after the visual detection task to a second client.
Further, the step of receiving a visual inspection task sent by a second client specifically includes: receiving visual inspection content sent by a second client and a model ID of a standardized model of the visual inspection content;
acquiring and calling a standardized model from a model library to execute the visual inspection task, specifically: and acquiring and calling a standardized model corresponding to the model ID from a model library to execute the visual inspection content.
Compared with the prior art, the invention has the beneficial effects that:
1) the multi-model cooperation system based on model arrangement is provided with an independent and universal model scheduling engine, and the model management background, the model library and the model scheduling engine are mutually independent, so that a client does not need to modify codes of an overall program when needing to replace or update a model, only needs to directly replace the model to the model library, or directly modifies the scheduling sequence of the model on a model arrangement tree, the model management background generates a new model scheduling relation table according to the modified part, the model scheduling engine only needs to perform model scheduling according to the relation table, and the whole system greatly improves the efficiency of a model scheduling process;
2) the model arrangement mode provided for the user by utilizing the model arrangement tree is intuitive and convenient to operate;
3) the input and output of the model are standardized, so that the same calling code can be used when different models are called, and Restful packaging is carried out on the model interface, thereby being beneficial to the remote calling of the model;
4) the invention adopts a mirror image file management mode to manage the model, and containerizes the model, thereby being beneficial to directly carrying out the model in an environment meeting the requirements when calling.
Drawings
Fig. 1 is a schematic structural diagram of a system according to embodiment 1 of the present invention.
FIG. 2 is a schematic structural view of a part model of embodiment 1 of the present invention.
Fig. 3 is a schematic flowchart of a process of uploading a standardized model by a first client in embodiment 1 of the present invention.
Detailed Description
The drawings are only for purposes of illustration and are not to be construed as limiting the invention. For a better understanding of the following embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product; it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
Example 1
As shown in fig. 1, the present embodiment provides a multi-model collaboration system based on model orchestration, where the system includes a first client 1, a second client 2, a model management background 3, a model library 4, and a model scheduling engine 5.
The first client 1 is used for constructing model arranging tree nodes, and the model library 4 is used for storing a plurality of standardized models; the first client 1 is configured to set a plurality of standardized models stored in the model library 4 on nodes of a model arrangement tree to construct a model arrangement tree, where the nodes of the arrangement tree are sequentially connected from front to back or connected in parallel, and the first client 1 may set the standardized models on the nodes of the model arrangement tree and arrange execution relationships between the standardized models according to relationships between the nodes of the model arrangement tree, where if two nodes where two standardized models are located are sequentially connected from front to back, the execution order of the two standardized models is sequentially executed from front to back. Preferably, the first client 1 arranges the model on the graphical interface, arranges the model by dragging the model to the node of the model arrangement tree, and arranges the model by simply dragging the model on the interface, so that the scheduler is visual and clear, and the operation is very simple and convenient.
The first client 1 is also used for binding a plurality of standardized models on all model arrangement tree nodes to a graphic processor, so that the standardized models are operated by adopting the bound graphic processor;
and the model management background 2 generates a model scheduling relation table according to the model arrangement tree constructed by the first client 1, the corresponding standardized models on all the nodes of the model arrangement tree and the graphics processor bound by the standardized models.
The model scheduling relation table comprises a tree structure of the model arrangement tree, a standardized model corresponding to each node in the model arrangement tree and a graphic processor bound by the model.
In the system of the present embodiment, the second client 2 is configured to send a visual inspection task to the model scheduling engine 5;
and the model scheduling engine 5 is used for calling a standardized model in the model library 4 to execute the visual detection task according to the model scheduling relation table in the model management background 3 after receiving the visual detection task, wherein the calling mode is to sequentially call the standardized model corresponding to each node on the scheduling tree according to the tree structure of the model scheduling tree in the model scheduling relation table. After the model completes the visual inspection task, the model scheduling engine 5 is used to return the completed result to the second client 2.
Alternatively, the "model" mentioned in the present embodiment may be a deep learning model, and the visual detection task may be a picture or an image, etc.
The standardized models mentioned in this embodiment are all models in which input and output of the models are standardized, and input and output of the initialization interface of the models are standardized separately, and the standardized specifications are only described below by way of example, but the way of standardizing the models in the system provided in this embodiment should not be limited to the following example descriptions:
the model initialization interface is standardized, and the input specification of the model can be: the input of any model only needs one parameter to be id, which represents the id of the model, and the type is long. The model initialization interface is used for loading the model into the GPU or the CPU in advance, so that the detection speed of the subsequent model is increased. The output of the model may be normalized as: the output format of any model is json type, and the output format of any model needs to comprise two basic fields, msg, which represent the initialization result of the model; code, a response status return value indicating an initialization request.
The visual inspection interface is standardized, and the input specification of the model can be as follows: only two parameters are required for the input of any one model: data, which represents photo data, a type byte array; id, model id, type long.
The output of the model may be normalized as: the output of any model needs to include two basic fields, timestamp, which represents the timestamp of the end of the model analysis; detections, representing the result of model classification or recognition. For the detection of model classification, probabletype is required to be included to respectively mark confidence and type for each classification result; for the model detection detections, positions representing the positions of the recognition objects are required to be included on the basis of the probableTypes.
Preferably, the first client 1 may set a component model on each model arrangement tree node, as shown in fig. 2, where the component model includes at least one standardized model, that is, at least one standardized model is set on the model arrangement tree node, and the component model is formed by combining specific standardized models by the client 1 according to a computation scenario.
Therefore, when the first client 1 sends the visual inspection task to the model scheduling engine 5, the visual inspection task includes the visual inspection content and the model ID for specifying the standardized model in the calling component model, and therefore, when the second client 2 sends the visual inspection task to the model scheduling engine 5, the specific process is as follows: the second client 2 sends the visual inspection content to the model scheduling engine 5, together with the model ID of the standardized model that needs to complete the visual inspection content. The model ID is a unique identification of the standardized model.
After receiving the visual inspection content and the model ID, the model scheduling engine 5 calls the standardized model in the model library correspondingly, and the specific calling process is as follows: and acquiring a tree structure of the model arrangement tree in the model scheduling relation table, and calling the component models on the nodes of the arrangement tree according to the sequence of the tree structure, wherein the component model on each node is provided with at least one standardized model, so that the model scheduling engine 5 accurately calls one of the standardized models on the nodes according to the model ID sent by the second client 2.
The concept of the component model is introduced, one component model is arranged on each model arrangement tree node, the component model comprises at least one standardized model, and one standardized model can be selected to be called when calling, so that a model arrangement mode is suitable for more calculation scenes.
Preferably, the model library 4 may be a mirror repository, as shown in fig. 3, and the first client 1 is further configured to upload a plurality of standardized models to the model management background 3; preferably, when uploading the standardized model, the first client 1 may upload model attributes corresponding to the model together, such as a model type, a model version number, a model type, a model name, a minimum hardware requirement for running, a photo type detected by the model, a model running environment, a photo feature, a detection allocation result, and the like.
After the model and the attribute are successfully uploaded by the first client 1, the model management background 3 is used for calling a Docker command to package the standardized model and the corresponding model attribute uploaded by the first client 1 into a mirror image file;
after the model management background 4 successfully stores the model and packages the model into a mirror image file, the first client 1 uploads a test task corresponding to the standardized model to the model management background 3; the test task is used for detecting whether the model can normally run after being uploaded.
After receiving the test task corresponding to the standardized model, the model management background 3 starts the image file of the standardized model to complete the test task, and returns the completion result of the test task to the first client 1;
after receiving the completion result of the test task, the first client 1 judges whether the standardized model successfully completes the test task according to the result; the model management background 3 determines whether to push the image file to an image warehouse according to the result determined by the first client 1: if the first client 1 judges that the standardized model successfully completes the test task, the model management background 2 pushes the image file to an image warehouse; otherwise, the model management background 2 will not push the image file, and the first client 1 may choose to upload the model again.
If the model mirror image is successfully pushed to the mirror image warehouse, the model scheduling engine 5 specifically pulls and starts the corresponding standardized model from the model mirror image warehouse when the model is called, the calling sequence is according to the tree structure of the model arrangement tree in the model scheduling relation table, and the mirror image file of the corresponding standardized model needs to be started according to the model ID sent by the first client 1 to complete the visual inspection task, and finally the completed result is returned to the first client 1.
The mirror image warehouse is used for managing the model to realize the sharing of the model, and meanwhile, the model is constructed into a container mirror image, so that the operation and maintenance difficulty of the model can be reduced, the deployment efficiency of the model at any end is improved, and the operation of the model is not influenced by the operation environment.
Preferably, the model management background 3 is further configured to perform Restful encapsulation on the received interface of the standardized model before calling a Docker command to package the standardized model and the model attribute uploaded by the first client 1 into an image file. The model management background 3 is responsible for performing Restful packaging on the interface of the standardized model, so that the standardized model can be called remotely, and the service of the standardized model is realized.
It should be understood that the above-mentioned embodiments of the present invention are only examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific embodiments of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention claims should be included in the protection scope of the present invention claims.
Claims (10)
1. A multi-model cooperation system based on model arrangement is characterized in that the system comprises a model library, a first client and a model management background:
the model base is used for storing a plurality of standardized models;
the first client is used for constructing model arrangement tree nodes and arranging the standardized models stored in the model library on the model arrangement tree nodes to form a model arrangement tree;
the first client is also used for binding the standardized model arranged on the model arrangement tree node to a graphic processor so that the standardized model runs by adopting the bound graphic processor;
the model management background is used for generating a model scheduling relation table according to the model arrangement tree formed by the first client, the standardized models arranged on the nodes of the model arrangement tree and the graphics processor bound by the standardized models.
2. The model orchestration based multi-model collaboration system of claim 1, further comprising a second client and a model scheduling engine,
the second client is also used for sending a visual detection task to the model scheduling engine;
and the model scheduling engine is used for acquiring and calling the standardized model from the model library to execute the visual detection task according to the model scheduling relation table generated by the model management background after receiving the visual detection task, and returning the execution result of the visual detection task to the second client.
3. The model orchestration based multi-model collaboration system of claim 2,
the second client is used for sending a visual inspection task to the model scheduling engine, and specifically comprises the following steps: the second client is used for sending visual detection content and the model ID of the standardized model required to be called for executing the visual detection content to a model scheduling engine;
the model scheduling engine is configured to, after receiving the visual inspection task, obtain and call the standardized model from the model library according to the model scheduling relationship table generated by the model management background to execute the visual inspection task, and specifically includes: and the model scheduling engine is used for acquiring and calling the standardized model corresponding to the model ID from the model library to execute the visual detection content according to the model scheduling relation table generated by the model management background after receiving the visual detection content and the model ID.
4. The model orchestration based multi-model collaboration system of claim 3, wherein the model repository is a model mirror repository;
the first client is also used for uploading a plurality of standardized models to the model management background;
the model management background is further used for generating an image file of the standardized model after receiving the standardized model uploaded by the first client and storing the image file to the image warehouse;
the model scheduling engine is configured to obtain and call the standardized model corresponding to the model ID from the model library according to the model scheduling relationship table generated by the model management background to execute the visual inspection content, and specifically includes: and the model scheduling engine pulls and starts the image file of the standardized model corresponding to the model ID from the model image warehouse according to the model scheduling relation table generated by the model management background so as to execute the visual detection content.
5. The model orchestration based multi-model collaboration system of claim 4,
the first client is also used for uploading the running test task corresponding to the image file to the model management background;
the model management background is further used for starting the generated image file to execute the running test task after receiving the running test task corresponding to the image file and generating the image file of the standardized model, and returning an execution result of the running test task to the first client;
the first client is further used for judging whether the image file successfully executes the running test task according to the execution result of the running test task after receiving the execution result of the running test task;
the model management background is further used for storing the image file to the model image warehouse when the first client determines that the image file successfully executes the operation test task.
6. The model orchestration based multi-model collaboration system according to any one of claims 4 or 5, wherein the model management background is further configured to perform Restful encapsulation on the received interface of the standardized model after receiving the standardized model uploaded by the first client and before generating an image file of the standardized model.
7. The model orchestration based multi-model collaboration system according to any of claims 2, 3, 4, and 5, wherein the first client and the second client are the same client.
8. A multi-model cooperation method based on model arrangement is characterized in that the method is applied to a first client side and comprises the following steps:
constructing model arrangement tree nodes, and arranging a plurality of standardized models in a model base on the model arrangement tree nodes to form a model arrangement tree; the standardized model is stored in the model library;
binding the standardized model provided on the model orchestration tree node to a graphics processor such that the standardized model runs with the bound graphics processor.
9. A multi-model cooperation method based on model arrangement is applied to a second client side, and comprises the following steps:
sending a visual detection task to a model scheduling engine so that the model scheduling engine acquires and calls a standardized model from a model library according to a model scheduling relation table to execute the visual detection task;
the model scheduling relation table is generated by a model management background according to a model arrangement tree formed by a first client, the standardized models set on the nodes of the model arrangement tree by the first client and a graphic processor bound to the standardized models by the first client; the standardized model is stored in the model library and is operated by the bound image processor;
and receiving the execution result of the visual detection task returned by the model scheduling engine.
10. A multi-model cooperation method based on model arrangement is characterized in that the method is applied to a model scheduling engine and comprises the following steps:
receiving a visual detection task sent by a second client;
obtaining a model scheduling relation table, wherein the model scheduling relation table is generated by a model management background according to a model arrangement tree formed by a first client, a standardized model of the first client, which is arranged on a node of the model arrangement tree, and a graphic processor of which the first client is bound to the standardized model, the standardized model is operated by adopting the bound image processor, and the standardized model is stored in a model library;
and acquiring and calling the standardized model from the model library to execute the visual detection task according to the model scheduling relation table, and returning an execution result of the visual detection task to the second client.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010419296.6A CN111752554B (en) | 2020-05-18 | 2020-05-18 | Multi-model cooperation system and method based on model arrangement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010419296.6A CN111752554B (en) | 2020-05-18 | 2020-05-18 | Multi-model cooperation system and method based on model arrangement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111752554A true CN111752554A (en) | 2020-10-09 |
| CN111752554B CN111752554B (en) | 2021-03-12 |
Family
ID=72673387
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010419296.6A Active CN111752554B (en) | 2020-05-18 | 2020-05-18 | Multi-model cooperation system and method based on model arrangement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111752554B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408639A (en) * | 2021-06-30 | 2021-09-17 | 山东新一代信息产业技术研究院有限公司 | Equipment identification model combination method and system for machine room inspection robot |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102193802A (en) * | 2011-05-19 | 2011-09-21 | 浙江大学 | Method for converting models with model subsets of same base class structure |
| US20140122028A1 (en) * | 2012-10-28 | 2014-05-01 | The Mathworks, Inc | Self-testing graphical component algorithm specification |
| US20140354630A1 (en) * | 2013-06-04 | 2014-12-04 | Electronics And Telecommunications Research Institute | Tree model and forest model generating method and apparatus |
| CN104657215A (en) * | 2013-11-19 | 2015-05-27 | 南京鼎盟科技有限公司 | Virtualization energy-saving system in Cloud computing |
| CN106453492A (en) * | 2016-08-30 | 2017-02-22 | 浙江大学 | Docker container cloud platform container scheduling method based on fuzzy mode recognition |
| CN107169022A (en) * | 2017-04-07 | 2017-09-15 | 广东精点数据科技股份有限公司 | A kind of system and method for realizing data warehouse automatic modeling |
| CN107391136A (en) * | 2017-07-21 | 2017-11-24 | 众安信息技术服务有限公司 | A kind of programing system and method based on streaming |
| CN110795219A (en) * | 2019-10-24 | 2020-02-14 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Resource scheduling method and system suitable for multiple computing frameworks |
| CN110941421A (en) * | 2019-11-29 | 2020-03-31 | 广西电网有限责任公司 | Development machine learning device and using method thereof |
| CN111083722A (en) * | 2019-04-15 | 2020-04-28 | 中兴通讯股份有限公司 | Model pushing method, model requesting method, model pushing device, model requesting device and storage medium |
-
2020
- 2020-05-18 CN CN202010419296.6A patent/CN111752554B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102193802A (en) * | 2011-05-19 | 2011-09-21 | 浙江大学 | Method for converting models with model subsets of same base class structure |
| US20140122028A1 (en) * | 2012-10-28 | 2014-05-01 | The Mathworks, Inc | Self-testing graphical component algorithm specification |
| US20140354630A1 (en) * | 2013-06-04 | 2014-12-04 | Electronics And Telecommunications Research Institute | Tree model and forest model generating method and apparatus |
| CN104657215A (en) * | 2013-11-19 | 2015-05-27 | 南京鼎盟科技有限公司 | Virtualization energy-saving system in Cloud computing |
| CN106453492A (en) * | 2016-08-30 | 2017-02-22 | 浙江大学 | Docker container cloud platform container scheduling method based on fuzzy mode recognition |
| CN107169022A (en) * | 2017-04-07 | 2017-09-15 | 广东精点数据科技股份有限公司 | A kind of system and method for realizing data warehouse automatic modeling |
| CN107391136A (en) * | 2017-07-21 | 2017-11-24 | 众安信息技术服务有限公司 | A kind of programing system and method based on streaming |
| CN111083722A (en) * | 2019-04-15 | 2020-04-28 | 中兴通讯股份有限公司 | Model pushing method, model requesting method, model pushing device, model requesting device and storage medium |
| CN110795219A (en) * | 2019-10-24 | 2020-02-14 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Resource scheduling method and system suitable for multiple computing frameworks |
| CN110941421A (en) * | 2019-11-29 | 2020-03-31 | 广西电网有限责任公司 | Development machine learning device and using method thereof |
Non-Patent Citations (2)
| Title |
|---|
| UN-BO WANG等: "A Machine Learning Framework for Resource Allocation Assisted by Cloud Computing", 《HTTPS://ARXIV.ORG/PDF/1712.05929.PDF 》 * |
| 谢乘胜: "面向深度学习应用的容器集群管理系统研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408639A (en) * | 2021-06-30 | 2021-09-17 | 山东新一代信息产业技术研究院有限公司 | Equipment identification model combination method and system for machine room inspection robot |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111752554B (en) | 2021-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2023071075A1 (en) | Method and system for constructing machine learning model automated production line | |
| WO2022135478A1 (en) | Task orchestration method and orchestration apparatus | |
| CN111506304A (en) | Assembly line construction method and system based on parameter configuration | |
| CN107908469B (en) | Task scheduling method and system | |
| WO2023004805A1 (en) | Workflow modeling implementation system and method, and storage medium | |
| CN111104103A (en) | Visualization method and system for software editing microservice | |
| KR100910336B1 (en) | A system and method for managing the business process model which mapped the logical process and the physical process model | |
| CN103530134B (en) | A kind of configurable software platform structure | |
| KR101326985B1 (en) | Method and apparatus for developing, distributing and executing object-wise dynamic compileless programs | |
| CN112596876A (en) | Task scheduling method, device and related equipment | |
| CN118245032B (en) | Attribute linkage engine method and system for customer relationship management | |
| CN111752554B (en) | Multi-model cooperation system and method based on model arrangement | |
| CN114912897A (en) | Workflow execution method, workflow arrangement method and electronic equipment | |
| CN117252559B (en) | Business process processing method, device, computer equipment and storage medium | |
| CN113448678A (en) | Application information generation method, deployment method, device, system and storage medium | |
| US20240265275A1 (en) | Device Deployment Method for AI Model, System, and Storage Medium | |
| CN112558930B (en) | Software generation system and method for container service | |
| CN117806654A (en) | Tekton-based custom cloud native DevOps pipeline system and method | |
| CN114115821A (en) | Application development method and platform, application deployment method and node, system and equipment | |
| CN113407174A (en) | Task scheduling method, device, equipment and storage medium | |
| CN116841758A (en) | Workflow task processing method, device, computer equipment and storage medium | |
| CN111752555B (en) | Business scene driven visual insight support system, client and method | |
| CN115115062B (en) | Machine learning model building method, related device and computer program product | |
| CN114296883B (en) | Light-load virtualized network experimental behavior simulator construction and scheduling method | |
| CN114490694A (en) | Business rule processing method and device, server and storage medium |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |