How ayuk007/mltunex works
Project type and purpose
• MLTuneX is an open-source AutoML library (Python package) for model training and hyperparameter tuning.
• It orchestrates data ingestion, model training, evaluation, hyperparameter tuning (via Optuna and GPT guidance), and model persistence.
File structure
• Top-level:
• setup.py, pyproject.toml, requirements.txt → packaging/configuration
• src/mltunex/… → core library code
• tests/… → unit and integration tests
• src/mltunex sub-packages:
• ai_handler → GPT-based prompt generation & metadata profiling
• config → global configuration loader
• data → ingestion, loading, splitting
• evaluate → model evaluation and metrics registry
• hyperparam_tuner → tuning orchestration, Optuna integration
• library_trainer → wrappers around sklearn trainers
• model_registry → registering & persisting best models
• trainer → generic training interface
• training_strategy → iteration & parallelization policies
• utils → shared helpers
README insights
• Entry point: MLTuneX(data, target_column, task_type).run(...)
• External dependencies: scikit-learn, Optuna, OpenAI API
• Workflow: train → evaluate → tune → save
Main components & their relationships
A. User / Client
B. CLI / Python API (main.py)
C. Config module
D. Data Pipeline
1. Ingestion
2. Loading
3. Splitting
E. Trainer
1. Base trainer
2. Library-specific trainers (sklearn)
F. Evaluation
1. Metrics registry
2. Evaluate_model
G. Hyperparameter Tuner
1. Tuning orchestration
2. Optuna tuner
3. AI Handler (GPT prompts & profiling)
H. Training Strategies (iteration / parallelization)
I. Model Registry (persist best model)
J. External services
1. Optuna
2. OpenAI API
K. Output
1. Result CSV
2. Saved model files
Relationships / Data Flow:
• User invokes API → main.py
• main → config loads settings
• Data pipeline reads CSV → produces train/test splits
• Trainer trains a set of models → passes to Evaluate
• Evaluate computes metrics → selects top models
• Hyperparam tuner orchestrates Optuna & optionally AI guidance → tunes top models
• Tuned models retrained via library_trainer → final evaluation
• Model registry saves best model → file system
• Results written to CSV
Architectural patterns & design principles
• Modular layering (data → training → evaluation → tuning → persistence)
• Strategy pattern (training_strategy for iteration/parallelization)
• Plugin-like extensibility (library_trainer, model_registry can support new libraries)
• Single responsibility per sub-package