Skip to content

BUG: merging posterior files can fail with tiny noise evidence discrepancies #1086

@mattpitkin

Description

@mattpitkin

Describe the bug

The bilbly_result script can be used to merge posterior results from parallel runs. When merging it checks for consistency between the runs by making sure that the null log likelihoods are the same. The absolute tolerance set for this comparison is 1e-8 and if not passed the script will fail. However, I have a few cases where numerical precision issues lead to files that are consistent being flagged as inconsistent and therefore not being merged, e.g.:

res1.log_noise_evidence
np.float64(68116786.40736821)
res2.log_noise_evidence
np.float64(68116786.4073682)
print(res1.log_noise_evidence - res2.log_noise_evidence)
1.4901161193847656e-08

Suggested solution (optional)

Increase the absolute tolerance for comparison to 1e-7.

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions