Titanic - PyTorch MLP (Kaggle)

A reproducible PyTorch MLP pipeline for the Kaggle Titanic dataset.

This project demonstrates:

• Modular feature engineering
• Clean training loop separation
• Ensemble inference
• Reproducible submission generation

Project Structure

• src/
  • model.py # MLP architecture definition
  • train.py # training loop (optimizer, loss, logging)
  • features.py # feature engineering (fit / transform pipeline)
• scripts/
  • make_submission.py # production entry point
• data/
  • raw/
    • train.csv
    • test.csv
• outputs/
  • submission/
    • submission.csv
• requirements.txt
• README.md

---

Setup

Install dependencies:

pip install -r requirements.txt

---

Data

• Download Titanic dataset from Kaggle: titanic

• Place the files here: data/raw/train.csv data/raw/test.csv

---

Run Submission Pipeline

• From project root: python scripts/make_submission.py

• Output file: outputs/submissions/submission.csv

---

Feature Pipeline

• Preprocessing is fitted on training data only.

• All statistics (median, quantile bins, encoders) are stored as artifacts.

• The same artifacts are reused for validation and test transformation.

• This ensures: No data leakage Stable feature space Reproducible inference

---

Model

• PyTorch MLP
• Configurable depth
• Dropout regularization
• Adam optimizer
• Ensemble across multiple random seeds

---

Reproducibility

• Randomness is controlled via: Numpy seed PyTorch seed Deterministic train / validation split

---

Notes

• The submission script is intentionally clean (no plotting or experiment logic).