mlflow-cynex

Cynex trajectory viewer integration for MLflow. Adds a "View in Cynex" button to MLflow's artifact browser for trajectory JSON files.

Installation

pip install mlflow-cynex
mlflow-cynex install

Then restart your MLflow server.

Usage

Logging Trajectories

The package provides a helper function to log trajectories to MLflow:

from mlflow_cynex import log_trajectory
import mlflow

mlflow.set_tracking_uri("http://localhost:5000")

with mlflow.start_run():
    trajectory = {
        "blue_agent_name": "PPO",
        "red_agent_name": "Meander",
        "episode": 0,
        "blue_actions": [...],
        "red_actions": [...],
        "metric_scores": [...],
        "network_topology": {...}
    }
    log_trajectory(trajectory)

The trajectory is automatically:

• Validated against the expected schema
• Saved to trajectories/<name>-trajectory.json in MLflow artifacts

Custom Naming

log_trajectory(trajectory, name="my-custom-name")
# Saved as: trajectories/my-custom-name.json

Validation Only

from mlflow_cynex import validate_trajectory, TrajectoryValidationError

try:
    validate_trajectory(data)
except TrajectoryValidationError as e:
    print(f"Invalid trajectory: {e}")

Trajectory Detection

The "View in Cynex" button appears for JSON files that:

• Are in a trajectories/ folder, OR
• Match *-trajectory.json or *_trajectory.json pattern

CLI Commands

mlflow-cynex install    # Install Cynex viewer into MLflow
mlflow-cynex uninstall  # Remove Cynex viewer from MLflow
mlflow-cynex status     # Check installation status

Development

Build from source:

cd cynex/mlflow-integration
./scripts/build.sh
pip install dist/mlflow_cynex-*.whl

License

MIT