canary — ML Model Monitoring

v1.2.5 · MIT · Python 3.9–3.12

CANARY

Drift and anomaly detection for production ML models. Cutting-edge performance in a wrapper, for free.

pip install canary-ml<br>copy

Keras/TensorFlow: pip install canary-ml[keras]

confidence_score · batch #2847

baseline

current

stable

PSI score

0.02

KS statistic

0.04

anomaly rate

0.8%

line to wrap any sklearn or Keras model

infrastructure required — runs fully local

from install to live monitoring dashboard

Data Drift Detection<br>KS test, PSI, chi-square per feature with configurable thresholds.

Anomaly Detection<br>Isolation Forest + z-score ensemble on inputs. KS test on prediction outputs.

Async Monitoring<br>Monitoring runs in a background thread. Your inference latency is unaffected.

Live Dashboard<br>Zero-dep HTML/JS dashboard. Ships with the package. No cloud account needed.

monitor_example.py

from canary_ml import ModelMonitor

monitor = ModelMonitor(<br>model=your_model,<br>reference_data=X_train,<br>alert_threshold=0.2,<br>log_path="./canary_logs"

# drop-in replacement — monitoring is a side effect<br>predictions = monitor.predict(X_new)

report = monitor.get_report()<br>print(report.psi_score, report.drift_detected, report.anomaly_rate)<br># 0.41 True 0.032

monitor.serve_dashboard(port=8501)