-
Notifications
You must be signed in to change notification settings - Fork 20
Guides | TANGO Architecture
giantlkh edited this page Oct 12, 2022
·
25 revisions
👉Note
- This page contains some wide rendered images. If you want to see the whole contents with wider page, please use Wide GitHub extension of Chrome.
- If you are experiencing the error on rendered images due to the low-bandwith Internet delay, please use F5 function key for refreshing current page.
TANGO (Taget Adaptive No Code Nueral Network Generation and Operation framework) is for novices on machine learning and deep learning area. It aims to provide easy and no-code based AI application build tool for novices who have some datasets but know how to utilize them for AI application. It is based on MSA (Mircro Service Architecture) technology, especially Docker containers.
In TANGO, there are two types of containers:
- A Main Container : Project Manager, which makes other containers start or stop as well as coordinates them to proceed workflow.
- One or More Member Containers: All containers except for Project Manager, which perform specific tasks in whole workflow.
Project Manager container (after here, simply Project Manager) provides project configuration and execution of workflow defined in the project configuration for login user. Here, project means the TANGO project, which provide No code based automatic neural network model generation, based on the user's deployment target HW and dataset, and deployment to the designated target HW with optimized neural network model for AI applications.
At TANGO project creation phase, users configure following information and save it.
- Dataset: custom dataset or public dataset (e.g., COCO dataset ).
- Target Device: on which the optimized neural network code deployed.
- User's customized workflow, user can use following task selectively
- NAS (Network Architecture Search) type used in
auto_nn** Backbone NAS or Neck NAS - use of visualizer for adding your own enhancement on the model generated from the
auto_nnphase
- NAS (Network Architecture Search) type used in
Upon completion of TANGO project creation, Project Manager coordinates the TANGO workflow; start, monitoring result of the member container execution for specific task and start next task if the previous task was successfully completed .
- Project Manager iterates over tasks one by one as defined in the workflow of TANGO project by starting corresponding member container.
Currently, TANGO consists of following member containers;
-
labelling: provides functionality to prepare user's own dataset annotation suitable for classification and detection tasks and COCO dataset import. -
base_model_select: suggests the basic neural network model suitable for user's dataset. -
auto_nn: generate an automatic neural network based on the neural network model selected frombase_model_selectand deploy target information defined project configuration.-
auto_nnconsists of Neck NAS (Network Architecture Search) and Backbone NAS. User select which NAS used forauto_nn
-
-
visualizer: provides editing functionally on the model fromauto_nn, if user select this feature in the workflow defined in project configuration. -
code_gen: generates template codes around the model, passed fromauto_nnorvisualizer, for use in deployment target. -
deploy: deploys the generated code incode_genon the specific target selected at project configuration-
deployconsists of ofcloud_deployxorondevice_deploy, which can be selected based on the user's choice on the target.
-
The following flowchart illustrates the whole process of TANGO from Project Creation to Deployment of Codes onto the Target Devices.
flowchart TB
subgraph P[TANGO Project Creation]
P1[Service Container Readiness Check]
P2[Project Configuration]
P1 --> P2
P2 --> P21[Dataset Select] & P22[Target Select] --> P3[Workflow Configuration]
P3 --> P4[Save Project]
end
S([Start]) --> P1
subgraph W[TANGO Workflow]
subgraph BMS[Base Model Select]
P4 --> W1[1. Base Model Select]
end
W1 --> C{Customize NN?}
C-->|Yes|V[Visualizer]
C -->|No| A1{NAS Type?}
V-->A1
subgraph A[2. AutoNN]
A1 -->|Neck NAS| A21[AutoNN based on Neck NAS]
A1 -->|Backbone NAS| A22[AutoNN based on Backbone NAS]
end
subgraph CG[3. Code Generation]
A21 & A22 --> W5[Code Generation]
end
subgraph D[4 Code Deployment]
W5--> W6{Deployment Target?}
W6 -->|Cloud Target| W71[Cloud Deploy]
W6 -->|OnDevice Target| W72[OnDevice Deploy]
end
end
W71 & W72 --> E([End])
Above flowchart was prepared with mermaid using Mermaid Online Editor
The following Sequence Diagram summarizes the overall control flow among containers within TANGO project.
sequenceDiagram
participant DC as Docker Compose
participant PM as Project Manager
participant LB as Labelling
participant BMS as Base Model Select
participant MV as Vis2Code
participant ANN as AutoNN
participant CG as Code Gen
participant CD as Cloud Deploy
participant OD as OnDevice Deploy
activate DC
%%TANGO Container Image Build
rect rgba(211, 211, 211, 0.3)
Note over DC : Build Container Images for TANGO
end
Note over DC, OD: TANGO Containers Instance Creation
rect rgba(211, 211, 211, 0.3)
Note right of DC : Start/Run Service Containers for TANGO
DC->>PM : docker run Project Manager
DC->>LB : docker run Labelling
DC->>BMS : docker run Base Model Selector
DC->>MV : docker run Vis2Code
DC->>ANN : docker run AutoNN
DC->>CG : docker run Code Gen
DC->>CD : docker run Cloud Deploy
DC->>OD : docker OnDevice Deploy
end
deactivate DC
Note over PM, OD: TANGO Project Configuration
rect rgba(127, 255, 212, 0.3)
activate PM
rect rgba(211, 211, 211, 0.3)
Note right of PM: Member Containers Health Check
PM ->> LB : status_request(user_id="", project_id="")
LB-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> MV : status_request(user_id="", project_id="")
MV-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> BMS : status_request(user_id="", project_id="")
BMS-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> ANN : status_request(user_id="", project_id="")
ANN-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> CG : status_request(user_id="", project_id="")
CG-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> CD : status_request(user_id="", project_id="")
CD-->>PM: HTTP response 200 OK with text/plain "ready"
PM ->> OD : status_request(user_id="", project_id="")
OD-->>PM: HTTP response 200 OK with text/plain "ready"
end
Note right of PM: Project Creation: Dataset and Target Configuration
rect rgba(211, 211, 211, 0.3)
opt Dataset Preparation
activate LB
Note right of LB: User defined Dataset labelling
deactivate LB
end
loop
Note over PM: Dataset Selection
PM->>PM:
Note over PM: Target Selection
PM->>PM:
end
end %% of rect rgba
end %% of Project Creation rect rgba(255, 95, 71, 0.7)
rect rgba(0, 255, 255, 0.3)
Note over PM, OD: TANGO Workflow based-on Project Settings
%% ====================
%% 1. Base Model Select
%% ====================
Note over PM, BMS: 1. Base Model Select
PM->>BMS: start(user_id=uid, prj_id=pid)
BMS-->>PM: HTTP response with "starting"
activate BMS
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
PM->>BMS: status_request(user_id=uid, project_id=pid)
BMS-->>PM: HTTP response 200 OK with text/plain "running"
end
end
alt Base Model Select Service Fail
%% Base Model Select Fail
rect rgba(255, 95, 71, 0.7)
break When Base Model Select Fails
Note over BMS: Base Model Select Fails
BMS->>PM:status_report(container_id="bms", user_id=uid, project_id=pid, result="failed")
end
end
Note right of PM: Workflow Fails at AutoNN
else Base Model Select Success
%% Base Model Select Success
rect rgba(173, 255,47, 0.7)
Note over BMS: Base Model Select Success
BMS->>PM:status_report(container_id="bms", user_id=uid, project_id=pid, result="success")
Note right of PM: Base Model Select Completes, Continue to directly AutoNN or via Visualizer
end
deactivate BMS
%% ======================================
%% Optional: Base Model Customization
%% ======================================
rect rgba(0, 206, 209, 0.3)
Note over PM, MV: Optional: Base Model Customization
PM->>MV: start(user_id=uid, prj_id=pid)
MV-->>PM: HTTP response with "starting"
rect rgba(211, 211, 211, 0.3)
activate MV
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
Note over MV: Base Model Customization
MV->>MV:
PM->>MV: status_request(user_id=uid, project_id=pid)
MV-->>PM: HTTP response 200 OK with text/plain "running"
end %% loop
end %% rect
end %% rect
rect rgba(255, 95, 71, 0.7)
Note over MV: Base Model Customization Fail
MV->>PM:status_report(container_id="vis2code", user_id=uid, project_id=pid, result="fail")
end
Note right of PM: Workflow Fail at Base Model Customization
rect rgba(173, 255,47, 0.7)
Note over MV: Base Model Customization Complete
MV->>PM:status_report(container_id="vis2code", user_id=uid, project_id=pid, result="success")
end
Note right of PM: Base Model Customization Completes, Continue to AutoNN
end %% rect for Optional: Base Model Customization
%% ====================
%% 2. AutoNN
%% ====================
Note over PM, ANN: 2. AutoNN
PM->>ANN: start(user_id=uid, prj_id=pid)
ANN-->>PM: HTTP response with "starting"
activate ANN
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
PM->>ANN: status_request(user_id=uid, project_id=pid)
ANN-->>PM: HTTP response 200 OK with text/plain "running"
end %% end of loop
end
alt AutoNN Fail
%% AutoNN Fail
rect rgba(255, 95, 71, 0.7)
break When AutoNN Fails
Note over ANN: AutoNN Fail
ANN->>PM:status_report(container_id="autonn", user_id=uid, project_id=pid, result="failed")
end
end
Note right of PM: Workflow Fails at AutoNN
else AutoNN Success
%% AutoNN Success
rect rgba(173, 255,47, 0.7)
Note over ANN: AutoNN Success
ANN->>PM:status_report(container_id="autonn", user_id=uid, project_id=pid, result="success")
deactivate ANN
end
Note right of PM: AutoNN Completes, Continue to Code Gen
%% ====================
%% 3. Code Gen
%% ====================
Note over PM, CG: 3. Code Gen
PM->>CG: start(user_id, prj_id)
CG-->>PM: HTTP response with "starting"
activate CG
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
PM->>CG: status_request(user_id=uid, project_id=pid)
CG-->>PM: HTTP response 200 OK with text/plain "running"
end
end
alt Code Gen Fails
%% Code Gen Fails
rect rgba(255, 95, 71, 0.7)
break When Code Gen Fails
Note over CG: Code Gen Fail
CG->>PM:status_report(container_id="code_gen", user_id=uid, project_id=pid, result="failed")
end
end
Note right of PM: Workflow Fails at Code Gen
else Code Gen Success
%% Code Gen Success
rect rgba(173, 255,47, 0.7)
Note over CG: Code Gen Success
CG->>PM:status_report(container_id="code_gen", user_id=uid, project_id=pid, result="success")
deactivate CG
end
Note right of PM: Code Gen Completes, Continue to Deployment
%% ====================
%% 4. Deployment
%% ====================
Note over PM, OD: 4. Deployment
alt Deployment to Cloud Target
%% ====================
%% 4.1 Cloud Deployment
%% ====================
Note over PM, CD: 4.1 Cloud Deployment
PM->>CD: start(user_id, prj_id)
CD-->>PM: HTTP response with "starting"
activate CD
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
PM->>CD: status_request(user_id=uid, project_id=pid)
CD-->>PM: HTTP response 200 OK with text/plain "running"
end %% end of loop
end
alt Cloud Deployment Fail
%% Cloud Deployment Fail
rect rgba(255, 95, 71, 0.7)
break When Cloud Deployment Fail
Note over CD: Cloud Deployment Fails
CD->>PM:status_report(container_id="cloud_deployment", user_id=uid, project_id=pid, result="failed")
end %% end of break
end %% of rect
Note right of PM: Workflow Fails at Cloud Deployment
else Cloud Deployment Success
%% Cloud Deployment Success
rect rgba(173, 255,47, 0.7)
Note over CD: Cloud Deployment Success
CD->>PM:status_report(container_id="cloud_deployment", user_id=uid, project_id=pid, result="success")
deactivate CD
end %% end of rec
end %% end of alt Cloud Deployment Fail/Success
Note right of PM: Complete Workflow
else Deployment to OnDevice Target
%% ====================
%% 4.2 OnDevice Deployment
%% ====================
Note over PM, OD: 4.2 Start OnDevice Deployment
PM->>OD: start(user_id, prj_id)
OD-->>PM: HTTP response with "starting"
activate OD
rect rgba(211, 211, 211, 0.3)
loop Every heartbeat period in Project Manager
PM->>OD: status_request(user_id=uid, project_id=pid)
OD-->>PM: HTTP response 200 OK with text/plain "running"
end %% end of loop
end
alt OnDevice Deployment Fail
%% OnDevice Deployment Fail
rect rgba(255, 95, 71, 0.7)
break When Cloud Deployment Fails
Note over OD: OnDevice Deployment Fail
OD->>PM:status_report(container_id="ondevice_deployment", user_id=uid, project_id=pid, result="failed")
end %% end of break
end %% end of rect
Note right of PM: Workflow Fails at OnDevice Deployment
else OnDevice Deployment Success
%% OnDevice Deployment Success
rect rgba(173, 255,47, 0.7)
Note over OD: OnDevice Deployment Successfully Completes!
OD->>PM:status_report(container_id="ondevice_deployment", user_id=uid, project_id=pid, result="success")
deactivate OD
end
Note right of PM: Complete Workflow
end %% end of alt else for OnDevice Deployment Fail/Success
end %% end of alt else for Cloud/On-Device Deployment
end %% end of alt else for Code Gen Fail/Success
end %% end of alt else for AutoNN Fail/Success
end %% end of alt else for BMS
end %% TANGO Workflow rect rgba(0, 255, 255, 0.3)
Above sequence diagram was prepared with mermaid using Mermaid Online Editor
👉Note
- This page might contain some wide rendered images. If you want to see the whole contents with wider page, please use Wide GitHub extension of Chrome.
- If you are experiencing the error on rendered images due to the low-bandwith Internet delay, please use F5 function key for refreshing current page.