[dirmeier]

Summary

Behaviour and API changes

• Prior-support rejection. The direct samplers (NPE, FMPE, NPSE, CMPE, AIO) take prior in sample_fn, and sample() forwards it. Draws outside the prior support are rejected, and DirectSampleInfo.acceptance_rate records the acceptance rate. Sequential proposals go through the same path.
• MCMC n_samples. It now counts the draws returned after warmup. With the NLE/NRE defaults, this returns 2000 draws per chain instead of 1000.
• NUTS warmup. Adaptation runs n_warmup steps, each chain keeps its own step size and mass matrix, and there is no second warmup.
• Kernel keyword arguments. make_sampler and the sample_with_* functions forward them: step_size for MALA and RMH, window-adaptation options for NUTS.
• sbc arguments. They are renamed to n_calibrations and n_samples.
• Batch norm removed. make_resnet, make_cnf and make_cm no longer take do_batch_norm or batch_norm_decay. The option crashed in a stateless hk.transform.
• Score models.
  • VE is the default SDE for NPSE and AIO, with separate sigma_min/sigma_max.
  • The VE base distribution is N(0, σ_max²).
  • Dividing the output by the marginal std is opt-in via scale_by_sigma.
• CMPE. Rewritten to follow Schmitt et al. (2023).
• Training.
  • Validation uses one key per run.
• Shape conventions (documented). Every distribution in the prior must be vector-valued, and the simulator must return (n, d), also for d = 1.

Bug fixes

• MCMC.
  • SBC crashed with NLE/NRE at default arguments.
• NPSE/AIO
  • log_prob failed for one-dimensional θ.
  • Setting head_size crashed the Simformer.
  • The time embedding was never used by the Simformer encoder.
• Summary networks
• ABC
  • SMC-ABC returned weighted particles without weights, started from a random ε, and crashed for a single parameter.
  • SABC reused a stale observation when the array was changed in place.
• Dead code removed.

Documentation

• New tutorial docs/notebooks/abc.ipynb covering:
• An "Implemented methods" table in the README and the docs.
• A caution box for the shape conventions.

[codecov]

Codecov Report

❌ Patch coverage is 96.09610% with 13 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.79%. Comparing base (6204646) to head (739fa10).

Files with missing lines	Patch %	Lines
sbijax/_src/experimental/nn/make_score_network.py	52.94%	8 Missing ⚠️
sbijax/_src/inference/summary/_compose.py	60.00%	2 Missing ⚠️
sbijax/_src/nn/make_flow.py	50.00%	2 Missing ⚠️
sbijax/_src/experimental/cmpe.py	96.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #93      +/-   ##
==========================================
+ Coverage   91.51%   92.79%   +1.28%     
==========================================
  Files         100      102       +2     
  Lines        3018     3111      +93     
==========================================
+ Hits         2762     2887     +125     
+ Misses        256      224      -32     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.