End-to-end applied ML pipelines (scikit-learn + PyTorch) with reproducible training, evaluation, and CLI inference.

Scope

This repository provides minimal, production-minded scaffolding for classical and simple deep learning tasks:

• deterministic data handling (seeded splits)
• config-driven training
• metric reporting
• artifactized model saving/loading
• CLI inference
• basic tests

The goal is to demonstrate clean applied ML engineering, not comprehensive MLOps or cloud deployment.

Structure

repo/
├─ README.md
├─ requirements.txt
├─ pyproject.toml   # optional
├─ src/
│  ├─ data/
│  ├─ models/
│  ├─ configs/
│  ├─ training/
│  ├─ evaluation/
│  ├─ tests/
│  └─ cli/
└─ notebooks/

Conventions:

• notebooks/ for EDA and prototyping only
• src/ contains reusable training/inference code
• configs/ contains YAML/JSON hyperparameter configs
• tests/ contains lightweight unit tests

Installation

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Usage

Training (example)

python -m src.training.train \
  --config configs/regression.yaml \
  --output artifacts/regression/

Evaluation (example)

Evaluation artifacts (metrics, plots, confusion matrices) are written to the output directory defined at train time.

Inference (example)

python -m src.cli.predict \
  --model artifacts/regression/model.joblib \
  --input samples/input.csv \
  --output predictions.csv

Reproducibility

• seeded splits (NumPy / Torch)
• pinned dependencies
• config-driven hyperparameters
• artifacts stored outside source tree

Tests

Lightweight tests verify data utilities, model initialization, and CLI inference. Run:

pytest -q

Status

Component	Status
Regression (SKL)	WIP
Classification	WIP
Torch baseline	WIP
CLI inference	WIP
Tests	WIP

Dataset Policy

No large datasets or sensitive data are committed. Small synthetic or toy CSVs may be used for tests.