Skip to content

README_en.md

Latest commit

 

History

History
200 lines (152 loc) · 8.32 KB

README_en.md

File metadata and controls

200 lines (152 loc) · 8.32 KB
Raw

AutoML Studio

简体中文 | English

An integrated workspace for multi-modal datasets, annotation, training, deployment, and AI-assisted production. In addition to the regular annotation flow, the platform now includes a full batch annotation pipeline: script ZIP upload, task orchestration, live run detail, failed-item retry, incremental execution, and result preview.

AutoML Studio

Capabilities

  • Dataset management for images, text, and multimodal conversation data
  • Annotation workbenches for detection, classification, segmentation, pose, LLM / MLLM, and DPO workflows
  • Training services, currently focused on YOLO detection and classification
  • Deployment services, currently focused on ONNX deployment and online inference
  • AI Pipeline for resource binding, template management, and capability orchestration
  • Batch annotation with tool ZIP upload, run detail, retry, incremental execution, and result preview

Architecture

flowchart LR
    browser[Browser] --> frontend[frontend_v2\nReact + Vite]
    frontend -->|/api| server[automl_server\nFastAPI]

    server --> mysql[(MySQL)]
    server --> minio[(MinIO)]
    server --> nacos[(Nacos)]
    server <--> rabbit[(RabbitMQ)]

    server -->|MQ train submit| rabbit
    rabbit -->|consume| trainer[model_trainer]
    trainer -->|MQ status log register| rabbit
    server -->|HTTP health cancel| trainer

    server -->|HTTP deploy infer| deploy[model_deploy]
    deploy -->|MQ deploy status| rabbit

    server -->|MQ RPC assist| rabbit
    rabbit -->|RPC worker| runtime[ai_pipeline_runtime]

    server -->|MQ batch execute| rabbit
    rabbit -->|consume| sandbox[ai_pipeline_sandbox]
    sandbox -->|MQ batch result| rabbit
Loading

Service Docs

Service Role Docs
frontend_v2 Frontend workspace for datasets, annotation, training, deployment, AI Pipeline, and batch annotation pages README
automl_server Main API and orchestration service for data, tasks, streams, and batch annotation persistence README / ARCHITECTURE
model_trainer Training service that consumes training jobs and sends status updates back README
model_deploy Deployment and inference service that manages runtime instances README
ai_pipeline_runtime AI capability runtime for image understanding, draft annotation, and white-overlay related flows README
ai_pipeline_sandbox Batch annotation script sandbox for ZIP execution, dependency install, and result callbacks README
model_training_runtime Experimental model training runtime; executes registered packages in one service with contract, data materialization, event, and result validation, and does not replace model_trainer yet README

Additional example:

Service Communication

Flow Main channel Notes
automl_server -> model_trainer MQ Training jobs are published to trainer.task.queue
model_trainer -> automl_server MQ Task status, task logs, and model registration are sent back over MQ
automl_server <-> model_trainer HTTP Used only for /health probing and /tasks/{task_id}/cancel
automl_server -> model_deploy HTTP Deploy, undeploy, deployment listing, deployment health, and inference are all direct HTTP calls
model_deploy -> automl_server MQ Successful deploy / undeploy is synced back through model.deployed and model.undeployed
automl_server <-> ai_pipeline_runtime MQ RPC Assist annotation and pipeline execution currently use RabbitMQ RPC rather than direct HTTP
automl_server -> ai_pipeline_sandbox MQ Batch chunks are published through pipeline.batch.execute
ai_pipeline_sandbox -> automl_server MQ Batch progress and results are returned through pipeline.batch.progress and pipeline.batch.result

The codebase still defines service.heartbeat as a message type, but trainer and deploy do not actively use a standalone MQ heartbeat right now. The status shown in the UI is currently based on on-demand /health checks from automl_server.

Current Batch Annotation Features

  • Tool layer: built-in scripts and uploaded ZIP script packages, with required root batch_script.json and optional root README.md
  • Management layer: tool list supports enable, disable, delete, and a detail drawer for editing name, description, and README.md
  • Run layer: run detail page supports SSE live progress, event timeline, paged results, original image preview, and YOLO overlay preview
  • Resume layer: continue queued runs, retry failed items only, and create incremental runs for newly added samples under the same configuration
  • Execution layer: ai_pipeline_sandbox creates isolated work directories, virtual environments, dependency caches, and structured error logs

Repository Layout

auto_ml/
├── frontend_v2/
├── automl_server/
├── model_trainer/
├── model_deploy/
├── ai_pipeline_runtime/
├── ai_pipeline_sandbox/
├── model_training_runtime/  # Experimental; direct in-service execution, does not replace model_trainer yet
├── mysql/
├── nacos/
├── readme/
├── docker-compose.yml
├── docker-compose.dev.yml
└── .env.example

Requirements

  • Docker 20.10+
  • Docker Compose v2
  • Node.js ^20.19.0 || >=22.12.0
  • Python 3.10+
  • pnpm >=10

Quick Start

cp .env.example .env
docker compose up -d

Default endpoints:

  • Frontend: http://localhost:3000
  • Main API: http://localhost:45678
  • Swagger UI: http://localhost:45678/swagger-ui
  • Trainer: http://localhost:8081
  • Deploy: http://localhost:8082
  • AI Runtime: http://localhost:8010
  • RabbitMQ console: http://localhost:15672
  • MinIO console: http://localhost:9010
  • Nacos: http://localhost:8848

Notes:

  • ai_pipeline_sandbox exposes its health endpoint on internal port 8011 and is not mapped to the host by default
  • The frontend container proxies /api to automl_server

Local Development

Start the core dependencies and the batch annotation sandbox first:

docker compose -f docker-compose.dev.yml up -d mysql rabbitmq minio minio-init nacos nacos-init ai-pipeline-sandbox

Frontend:

cd frontend_v2
pnpm install
pnpm dev

Main service:

cd automl_server
pip install -r requirements.txt
python run.py

Start the other services on demand as described in their own READMEs.

Configuration

  • Runtime configuration is primarily driven by Nacos AUTO_ML_CONFIG, with nacos/automl-config.yaml as the source file
  • Root .env.example is mainly for docker-compose.yml bootstrap values
  • automl_server/.env.example is for standalone backend runs
  • frontend_v2/.env.example is for frontend development and builds
  • Batch annotation highlights:
    • ai-pipeline-batch.secret_key: required to encrypt and decrypt script secret parameters
    • ai-pipeline-sandbox: controls timeout, memory, process count, venv paths, pip cache, and pip indexes
    • rabbitmq: includes pipeline_batch_execute, pipeline_batch_progress, and pipeline_batch_result queues and routing keys

Database and Migrations

  • Initial schema: mysql/init/01_init_automl.sql
  • Incremental migrations: mysql/migrations/*.sql

Recent batch annotation related migrations include:

  • 20260728_create_ai_pipeline_batch_annotation.sql
  • 20260728_add_user_batch_scripts.sql
  • 20260729_add_batch_annotation_incremental_runs.sql
  • 20260730_add_batch_script_readme.sql

Acknowledgments

This project is built on top of the following open-source projects: