How shalinvachheta017/masterarbeit_mlops works
This is an offline-first MLOps pipeline for wearable-sensor human activity recognition. Its principal workflow turns accelerometer and gyroscope recordings into validated, transformed windows, predictions, evaluation and monitoring evidence, then a trigger decision. Retraining, registration and baseline updates are a separate approval-gated path; the README says stages 8–10 require human review and are not automatic. The production pipeline orchestrates the runtime stages and passes stage artifacts. The API is an additional inference entry point. Unsampled components are included where documented, but their internal calls are not inferred.
Main entry points: production_pipeline.py Production orchestrator · model_inference.py Model inference · data_transformation.py Window transformation · post_inference_monitoring.py Post-inference monitoring
Components 23
Inference pipeline
- Production orchestrator
src/pipeline/production_pipeline.py - Sensor ingestion
src/components/data_ingestion.py - Data validation
src/components/data_validation.py - Window transformation
src/components/data_transformation.py - Model inference
src/components/model_inference.py - Prediction evaluation
src/components/model_evaluation.py - Post-inference monitoring
src/components/post_inference_monitoring.py - Drift analysis
src/monitoring/drift_monitor.py - Trigger evaluation
src/components/trigger_evaluation.py
Model governance
- Model retraining
src/components/model_retraining.py - Model registration
src/components/model_registration.py - Model rollback
src/model_rollback.py - Baseline update
src/components/baseline_update.py - Calibration uncertainty
src/components/calibration_uncertainty.py
Entry points
- Pipeline CLI
run_pipeline.py - Inference API
src/api/app.py
Shared internals
- Stage artifacts
src/entity/artifact_entity.py - Scaler integrity
src/utils/scaler.py - Activity label mapping
src/utils/label_mapping.py - MLflow tracking
src/mlflow_tracking.py
Other components
- Sensor recordings
- Pipeline operator
- Human reviewer
Connections 24
- Pipeline operator → Pipeline CLI: starts run
README.md - Pipeline CLI → Production orchestrator: runs pipeline
run_pipeline.py - Sensor recordings → Sensor ingestion: provides recordings
README.md - Production orchestrator → Sensor ingestion: runs stage
src/pipeline/production_pipeline.py - Sensor ingestion → Data validation: fused data
README.md - Data validation → Window transformation: validated data
README.md - Window transformation → Model inference: windowed input
README.md - Model inference → Prediction evaluation: predictions
README.md - Prediction evaluation → Post-inference monitoring: evaluation evidence
README.md - Post-inference monitoring → Drift analysis: analyzes drift
- Post-inference monitoring → Trigger evaluation: monitoring evidence
README.md - Trigger evaluation → Human reviewer: requests review
README.md - Human reviewer → Model retraining: approves retraining
README.md - Model retraining → Model registration: trained model
README.md - Model registration → Baseline update: registered model
README.md - Model registration → Model rollback: rollback support
src/components/model_registration.py - Pipeline operator → Inference API: uploads data
README.md - Inference API → Model inference: requests inference
README.md - Production orchestrator → Calibration uncertainty: optional stage
src/pipeline/production_pipeline.py - Window transformation → Scaler integrity: uses scaler
src/components/data_transformation.py - Model inference → Activity label mapping: uses mappings
src/components/model_inference.py - Inference API → Activity label mapping: uses mappings
src/api/app.py - Production orchestrator → Stage artifacts: passes artifacts
src/pipeline/production_pipeline.py - Model retraining → MLflow tracking: tracks training