sbi_snpe_training_loop

Imported from yzj623/pplab (basicagent-skill/skill_library/paperbench/sequential-neural-score-estimation/skill/sbi_toolkit/sbi_snpe_training_loop/SKILL.md). Install upstream with npx skills add yzj623/pplab --skill sbi_snpe_training_loop. Copyright stays with the author.

SBI SNPE Training Loop

Use this skill to train or emulate a small conditional neural posterior estimator from simulated (theta, x) pairs and query it at an observation.

Do not use it for SNLE/SNRE objectives or for posterior validation alone.

Inputs

• Simulated theta and x pairs.
• Prior bounds or prior metadata.
• Observation x_o.
• Bounded training hyperparameters and seed.

Outputs

• Posterior samples for x_o.
• Training diagnostics and mechanism checks.
• Declaration of full-package or reduced/proxy execution.

Workflow

1. Validate that theta and x are aligned and finite.
2. Fit a conditional posterior estimator that depends on x.
3. Sample from the estimated posterior conditioned on x_o.
4. Record loss/proxy diagnostics and sample summaries.
5. Mark reduced implementations as proxy recovery.

Validation

Run:

python scripts/snpe_proxy.py --num-simulations 256 --observation 1.25
python tests/test_snpe_proxy.py

Limitations

• The included script is a reduced Gaussian SNPE-style proxy for fast recovery, not the full sbi package.
• It must be declared as proxy recovery in soft mode.