Skip to content

Bring your own likelihood: external trainers

You trained a likelihood surrogate outside the HSSM ecosystem's LAN pipeline — in sbi or BayesFlow. This page routes you to the right integration path.

The routes

Route Source Mechanism When to use Tutorial
NRE → ONNX sbi lanfactory.onnx.transform_sbi_to_onnx → loglik="file.onnx" Portable, shareable artifact; ratio estimators sbi NRE integration
NLE → ONNX BayesFlow lanfactory.onnx.transform_bayesflow_to_onnx → loglik="file.onnx" Portable artifact; flow-based density estimators BayesFlow NLE integration
LRE → JAX callable BayesFlow in-memory JAX function → loglik=<callable> Fast iteration during model development; no export step BayesFlow LRE integration

Not yet supported: MNLE-style mixed discrete/continuous observations and NSF-based flows (blocked on SearchSorted ONNX support).

The shared workflow

All three tutorials follow the same skeleton, so you can read any one of them and transfer the pattern:

  1. Simulate training data from a known ground truth (ssm-simulators or hssm.simulate_data).
  2. Train the surrogate in its home library.
  3. Export — for the ONNX routes, via LANfactory's exporters, which enforce the ONNX likelihood contract by construction (see LANfactory's sbi and BayesFlow export guides for the framework-specific constraints).
  4. Load into HSSM — hssm.HSSM(loglik=..., loglik_kind="approx_differentiable"); HSSM does not need to know which framework trained the surrogate.
  5. Validate — for the DDM you can compare against HSSM's analytical likelihood as a gold standard; in general, check recovery on simulated data before trusting the surrogate on real data.

See also