1350 leakage correction breaks consistency of log prob vs log prob batched

[manuelgloeckler]

This should fix the issue #1350.

Changes

• accept_reject_sample now returns a tensor of acceptance_rates.
• Only change was needed for the restriction estimator. Which now reduces this tensor to a float.
• Test added and extended to bounded priors to detect inconsistencies with leakage correction methods.

[janfb]

Looks good, thanks for fixing this! 🙏

But where was the bug then here? Was it because the min_acceptance_rate was returned, which was the min over the batch of xs in case of log_prob_batched?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 78.42%. Comparing base (d3f22b5) to head (f512659).
Report is 7 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1355       +/-   ##
===========================================
- Coverage   89.40%   78.42%   -10.99%     
===========================================
  Files         118      118               
  Lines        8715     8749       +34     
===========================================
- Hits         7792     6861      -931     
- Misses        923     1888      +965     

Flag	Coverage Δ
unittests	78.42% <100.00%> (-10.99%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
sbi/samplers/rejection/rejection.py	88.00% <100.00%> (ø)
sbi/utils/restriction_estimator.py	76.31% <100.00%> (-8.59%)	⬇️

... and 32 files with indirect coverage changes

[manuelgloeckler]

But where was the bug then here? Was it because the min_acceptance_rate was returned, which was the min over the batch of xs in case of log_prob_batched?

The bug was that different xs in batched sampling lead to different posteriors. But the leakage correction scaled every "posterior" with the "min_acceptance_rate" i.e. the posterior which leaks most. But different posteriors, have different leakage.