Code for the paper "Frozen Priors, Fluid Forecasts", accepted at ICLR 2026.
conda env create -f environment.yml
conda activate predictive_flows| Script | Description |
|---|---|
Exp_A.py |
Two-Moons (2D) — RealNVP + Martingale Posterior |
Exp_B.py |
GPT-2 text generation — Martingale Posterior vs. Bootstrap |
Exp_B_Baselines.py |
GPT-2 text generation — Bayesian Bootstrap, DWS, Jackknife |
Exp_C_ID.py |
CIFAR-10 (in-distribution) — DDPM + Martingale Posterior vs. Bootstrap |
Exp_C_ID_Baselines.py |
CIFAR-10 (in-distribution) — Bayesian Bootstrap, DWS, Jackknife |
Exp_C_OOD.py |
SVHN (out-of-distribution) — DDPM + Martingale Posterior vs. Bootstrap |
Exp_C_OOD_Baselines.py |
SVHN (out-of-distribution) — Bayesian Bootstrap, DWS, Jackknife |
Each script saves results to a timestamped JSON file.
Python 3.10, PyTorch ≥ 2.6. Experiments B and C require a CUDA GPU; Experiment A runs on CPU.