[FIX] ASR: make nonlinear_eigenspace reproducible - #118
Merged
Conversation
This was referenced Jul 13, 2026
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #118 +/- ##
==========================================
- Coverage 83.10% 83.10% -0.01%
==========================================
Files 25 25
Lines 2818 2817 -1
==========================================
- Hits 2342 2341 -1
Misses 476 476 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
nonlinear_eigenspace drew its optimization starting point from an unseeded module-level Generator, so the riemann ASR path produced slightly different results on every call. The starting point only affects convergence, not the resulting eigenspace, so seed it with a fixed value. This makes the riemann ASR output reproducible run-to-run and stabilizes the incremental-vs-bulk equivalence check in test_asr_class. Adds a regression test asserting repeated calls return identical results.
sappelhoff
force-pushed
the
fix/asr-riemann-reproducible
branch
from
July 13, 2026 17:07
74d73d5 to
fa2ad8d
Compare
This was referenced Jul 13, 2026
Closed
nbara
approved these changes
Jul 16, 2026
nbara
added a commit
that referenced
this pull request
Jul 18, 2026
…ut guards (#121) A small bundle of documentation and defensive-hardening fixes: - **`fit_eeg_distribution` docstring:** the `min_clean_fraction` / `max_dropout_fraction` default values were transposed relative to the signature (`0.25` / `0.1`); corrected. Also `X` was described as a 2-D `(n_channels, n_samples)` array, but the function operates on a **1-D** vector of amplitude values (it does `np.sort(X); n=len(X)`) — description fixed. - **`asr_calibrate` non-finite guard:** a `NaN`/`Inf` in the calibration data previously propagated silently through the filter into `M`/`T` (an all-`NaN` threshold, no error). Non-finite samples are now zeroed before filtering so calibration degrades gracefully. - **`asr_calibrate` insufficient-data guards:** raise a clear `ValueError` when the calibration data is too short to form one analysis window, or yields fewer than two threshold-estimation windows — instead of an obscure downstream error. ### Testing Adds tests that non-finite calibration input yields finite `M`/`T`, and that too-short calibration raises a clear `ValueError`. Full `tests/test_asr.py` + `tests/test_cov.py` pass; `ruff` clean. --- **Note:** includes the `nonlinear_eigenspace` reproducibility fix from #118 so the riemann CI test is stable regardless of what this PR changes. Once #118 merges first, that commit drops out on rebase. Co-authored-by: Nicolas Barascud <10333715+nbara@users.noreply.github.com>
nbara
pushed a commit
that referenced
this pull request
Jul 20, 2026
…120) `fit_eeg_distribution`'s vectorized grid-search histogram divides by a per-column baseline-relative value that is **zero** when a run of samples is tied (e.g. a flatlined / disconnected electrode): `cols = nbins / newX[mcurr]` → `inf`, then `H = newX[:m] * cols` → `nan` (0·inf). Casting `nan` to int yields `INT64_MIN`, and `np.bincount` then raises `ValueError: 'list' argument must have no negative elements`. A flat segment is exactly the dropout case `max_dropout_fraction` is meant to tolerate, so the fit should degrade gracefully rather than crash. This makes the histogram binning ignore non-finite entries (count them in no bin), so a degenerate grid column just yields a poor — but finite — objective and the grid search continues. Results on finite input are unchanged. The crash is reachable through the primary public entry point: `ASR().fit(X)` with a single dead channel currently raises `ValueError`. ### Testing Adds three tests: the direct `fit_eeg_distribution` crash on a tied/flat vector; `ASR().fit()` with a dead channel; and a guard that the result on clean data is numerically identical (same `mu, sig, alpha, beta`). Full `tests/test_asr.py` + `tests/test_cov.py` pass; `ruff` clean. --- **Note:** includes the `nonlinear_eigenspace` reproducibility fix from #118 so the riemann CI test (`test_asr_class[*-riemann]`) is stable regardless of what this PR changes. Once #118 merges first, that commit drops out on rebase.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
nonlinear_eigenspacedrew its optimization starting point from an unseeded module-levelnp.random.default_rng(), so the riemann ASR path produced slightly different results on every call. Because the manifold optimizer converges to the same eigenspace regardless of the starting point, this only injected run-to-run numerical noise — but that noise makes the riemann results non-reproducible and destabilizes the incremental-vs-bulk equivalence check intest_asr_class(its< 6µVtolerance can be exceeded depending on dependency versions and the covariance regime).Seed the starting point with a fixed value so the riemann path is reproducible run-to-run. The mathematical result is unchanged; only the previously-random initialization is now deterministic.
Testing
Adds
test_nonlinear_eigenspace_reproducibleasserting repeated calls return identical results (fails on the unseeded code, passes after). Fulltests/test_cov.pyandtests/test_asr.pypass;ruffclean.