RA2_inverse_model_CMC

RA2 Inverse Model (CMC): End‑to‑End Reproduction

This repository reproduces the full experimental pipeline for the CMC inverse modeling project:

• Forward-model sanity checks (QC)
• Synthetic dataset generation (CMC simulator + regime rejection)
• Feature/token extraction (ERP/TFR/Hybrid) + fixed train/val/test splits
• Training inverse models (Transformer ensembles + BiLSTM baselines)
• Evaluation and diagnostics (calibration/reliability, SBC, PPC, metrics)
• Ablations:
  • Diagonal posterior head vs full covariance
  • Transformer WITHOUT per‑parameter tokens (“noparamtoken”)
• SBI baseline (SNPE) + mismatch robustness + PPC for SNPE

All commands below are designed to be run from the repo root.

---

Repository layout (what’s source vs generated)

Source (tracked)

• data/ dataset generation + preprocessing code
• models/ training code + architectures
• eval/ evaluation + plotting/diagnostics
• sim/ CMC forward simulator
• scripts/ convenience/NeurIPS/SBI scripts
• config/ config templates

Generated (do not commit)

• data/synthetic_cmc_dataset*.h5
• data_out/ (token features + splits)
• models_out/ (trained models + scalers + logs)
• plots/ (QC + evaluation figures + summary CSV/JSON)
• results/ (SBI/SNPE baseline outputs)

---