Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Xander44-4/traffic_reducer/llms.txt

Use this file to discover all available pages before exploring further.

The default app.run(debug=True) configuration in app.py is intended for local development only. It enables the interactive debugger, exposes the Werkzeug reloader (disabled here via use_reloader=False), and binds only to 127.0.0.1, making the server unreachable from other machines. This page covers the changes needed to run Traffic Reducer reliably and securely in a production environment.

​
Disable debug mode

Edit the final block in app.py to turn off debug mode and bind to all network interfaces: 
# In app.py — change:
app.run(debug=True, port=5000, use_reloader=False)

# To:
app.run(debug=False, port=5000, use_reloader=False, host='0.0.0.0')
Setting host='0.0.0.0' makes the server listen on all available interfaces so that other machines on the network can reach the dashboard and MJPEG stream.

​
Run with a production WSGI server

Flask’s built-in server is single-threaded and not hardened for production traffic. Replace it with a production-grade WSGI server: 
# Windows — Waitress
pip install waitress
waitress-serve --port=5000 traffic_app.app:app

# Linux / macOS — Gunicorn
pip install gunicorn
gunicorn -w 1 -b 0.0.0.0:5000 traffic_app.app:app
Always use exactly one worker process (-w 1 for Gunicorn). TrafficCamera maintains a global capture thread, an in-process YOLO executor, and a shared traffic_state dictionary. These cannot be safely shared or duplicated across multiple worker processes. Launching more than one worker will result in multiple competing capture threads and inconsistent state.

​
GPU acceleration

If the production machine has a CUDA-capable GPU, install the GPU-enabled build of PyTorch before installing Ultralytics. This allows YOLOv8 to run inference on the GPU, significantly reducing per-frame latency: 
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install ultralytics
Replace cu121 with the CUDA version matching your driver (e.g., cu118 for CUDA 11.8). Ultralytics will automatically use the GPU when PyTorch detects a CUDA device.

​
Fix MODEL_PATH

Before deploying, update the hardcoded MODEL_PATH in app.py to an absolute path valid on the production machine: 
MODEL_PATH = "/opt/traffic_reducer/traffic_reducer_dataset/modelo_entrenado/modelo_semaforo_ia.pkl"
Ensure the file exists at that path and is readable by the process user. If the model cannot be loaded, the /predict endpoint will return a 500 error for any request that does not use live_mode.

​
Run as a systemd service (Linux)

To have Traffic Reducer start automatically on boot and restart on failure, create a systemd unit file: 
[Unit]
Description=Traffic Reducer
After=network.target

[Service]
User=www-data
WorkingDirectory=/opt/traffic_reducer
ExecStart=/opt/traffic_reducer/venv/bin/python traffic_app/app.py
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target
Save the file to /etc/systemd/system/traffic-reducer.service, then enable and start it: 
sudo systemctl daemon-reload
sudo systemctl enable traffic-reducer
sudo systemctl start traffic-reducer
Check the service status and logs with: 
sudo systemctl status traffic-reducer
sudo journalctl -u traffic-reducer -f

​
Network access

By default, without the host='0.0.0.0' change described above, the server binds only to 127.0.0.1:5000 and is not reachable from other machines. After setting host='0.0.0.0', any machine on the same network can access the dashboard at http://<server-ip>:5000.
The MJPEG stream served at /video_feed sends a continuous sequence of JPEG frames encoded at quality 78–82. This is bandwidth-intensive, particularly on slow or wireless networks. If multiple clients need to view the stream simultaneously, consider proxying through nginx with buffering enabled to absorb burst traffic and reduce the load on the Flask/WSGI process.

Build docs developers (and LLMs) love

Get started for freeTalk to us