How lettezuart/xgboost-modelgpu works
Project Type and Purpose
• This is a data-science/ML pipeline project for time-series prediction of patient glucose levels (not a web app or microservice).
• It ingests raw CSV data, applies a Kalman filter, trains XGBoost models for two horizons (30 min, 60 min), then runs a final validation to produce metrics and plots.
High-Level Components
a. Data Sources
• Raw CSV files under Data/Raw/<patient_id>
b. Data Ingestion & Preprocessing
• utils/data_loader.py loads CSVs
• utils/data_processing.py & utils/features.py apply data cleaning, feature extraction
• utils/filter.py (Kalman filter only)
c. Dataset Abstraction
• utils/datasets.py constructs train/test splits
d. Offline Model Training
• models/training.py orchestrates training loops per patient & horizon
• models/offline_model.py defines XGBoost model workflows
• Outputs: model artifacts (model_param.json, scaler.pkl, metadata.json) written to Trained_model_outputs
e. Final Validation
• utils/final_validation.py loads saved models + scalers, predicts test sets
f. Evaluation
• utils/evaluation.py computes RMSE, MAE, R²
• utils/clarke.py computes Clarke error zones
g. Reporting & Visualization
• Training_metrics/*.txt and Validation_metrics/*.txt summary tables
• Training_plots/ and Validation_plots/ PNG graphs
h. Orchestration
• main.py invokes the full pipeline in sequence
Relationships and Data Flow
• → Dataset Abstraction → Offline Model Training
• → Trained_model_outputs (artifact store)
• → Final Validation → Evaluation → Reporting & Visualization
• main.py spans all stages, controlling the sequence
Architectural Patterns and Principles
• Pipeline pattern: discrete stages with well-defined inputs/outputs
• Modular utilities: separate filter, loader, processing, evaluation
• Separation of concerns: model definition vs. training vs. validation
• Artifact management: separate storage per patient/horizon
Technologies & Frameworks to Annotate
• Python 3.x
• pandas, numpy for data handling
• XGBoost (models/offline_model.py)
• joblib or pickle (scaler.pkl)
• matplotlib for plots
• tabulate for text tables