Neural Causal Regularization

NCR is a light‑weight Python library that implements the risk‑gap penalty for training deep models that generalise across distribution shifts. It ships with:

• 📦 train_ncr — 30‑line training loop for any torch.nn.Module
• 🔌 risk_gap & risk_variance penalties (NCR & REx)
• 🧰 helpers: accuracy, tune_lambda, toy ColoredMNIST generator

Why NCR?

Standard ERM fits average performance and often memorises spurious shortcuts. NCR adds an ℓ₁ penalty on the risk difference |R1 − R2| between environments, nudging the network toward features that are stable — typically the causal ones.

Get around

Use the sidebar (left) to jump to:

• Installation: pip / conda commands.
• Getting Started: 20‑line Python example on ColoredMNIST.
• Theory: formal objective & link to the 2025 paper draft.
• API Reference: every public function with signatures.