Skip to content

File Structure (Python Inference Service)

Overall Project Structure

rl-inference-service/
├── app/
│   ├── main.py                              # FastAPI application (MAPPO model inference)
│   ├── templates/
│   │   └── index.html                       # Landing page with navigation links
│   └── static/
│       ├── logo.png                         # Project logo
│       └── favicon.ico                      # Browser tab icon
├── trained_models/
│   └── agent.th                             # Trained MAPPO model weights (PyTorch state_dict)
├── Dockerfile                               # Python 3.9-slim base, Uvicorn server
├── requirements.txt                         # FastAPI, Uvicorn, PyTorch, Numpy, Pydantic, Jinja2
├── .env.example                             # Template: MAPPO_AGENT_PATH, API_HOST, API_PORT
├── .gitignore
├── README.md                                # Setup and deployment guide
└── docker-compose.override.yml              # Local development overrides (optional)

app/ Directory Structure

app/
├── main.py                                  # FastAPI instance, route handlers
│   ├── startup events (load model, initialize GRU states)
│   ├── POST /predict_action (junction_id, obs_data → action, confidence)
│   ├── POST /reset_hidden (reset all GRU hidden states)
│   ├── GET /health (service status, model_loaded)
│   ├── GET /model_info (model architecture, junctions, action masks)
│   ├── GET / (landing page)
│   └── GET /docs (Swagger UI)
├── templates/
│   └── index.html (static HTML with links)
└── static/
    └── assets (logo, favicon, etc.)

Key Files

  • main.py — All application logic in one file:
  • FastAPI app initialization
  • MAPPO RNNAgent loading from MAPPO_AGENT_PATH
  • Per-junction GRU hidden state management
  • 5 route handlers

  • Pydantic request/response models

  • trained_models/agent.th — PyTorch model file (~1-2 MB)

  • Loaded at startup via torch.load()

  • Contains RNNAgent weights for all 5 junctions

  • Dockerfile — Production image build

  • Base: python:3.9-slim
  • Installs requirements
  • Runs: uvicorn main:app --host 0.0.0.0 --port $API_PORT

Environment Variables

Variable Default Purpose
MAPPO_AGENT_PATH /app/trained_models/agent.th Path to model weights
API_HOST 0.0.0.0 Bind address
API_PORT 8000 Service port
API_RELOAD false Auto-reload on code changes