Dynamic β-NLL Uncertainty Estimation Framework

Modular PyTorch research framework to study dynamic β scheduling for heteroscedastic regression.

Getting Started

• Create env: python -m venv .venv && source .venv/bin/activate
• Install deps: pip install -r requirements.txt (add Hydra, PyTorch, WandB, PyTest, Ruff/Black)
• Run training: python train.py
• Evaluate: python eval.py checkpoint=path/to/ckpt.pt

Key Files

• configs/ Hydra configs (datasets, models, experiments)
• src/modules/loss.py GaussianLogLikelihoodLoss (do not change math) and BetaScheduler
• src/models/mlp.py backbones outputting mean and log_variance
• src/data/base_dataset.py toy long-tail dataset and loaders
• train.py dynamic β training loop with logging and grad norms
• eval.py calibration and NLL evaluation