Imports#1
Merged
AllenWLynch merged 13 commits intomainfrom Apr 28, 2026
Merged
Conversation
- sbs: max-subtract log_locus/log_context before exp in _get_exp_offsets_k_c and restore the shift on the scalar normalizer in log-space; fall back to the previous normalizer if the slice is still degenerate - theta_model: clip raw predictions to [-30, 30] at the log_locus_distribution write point so exp() stays within float32 - context_model: clip _get_log_context_distribution to [-30, 30] - optim: finite-check normalizers after every get_exp_offsets_dict call and re-enable init_normalizers so epoch 1 doesn't start from zero
…om saved state Pickling a TopographyModel was failing with "_pickle.PicklingError: Can't pickle <class 'mutopia.gtensor.xarr_extensions.FetchSample'>: it's not the same object as mutopia.gtensor.xarr_extensions.FetchSample". Two compounding issues: 1. xarr_extensions.py defined four classes named FetchSample (for list_samples, mutate, fetch_sample, iter_samples) and two named AsCSR (for ascsr and asdense). Only the last definition is bound to the module name, so pickle's class-identity check failed for any of the shadowed classes once xarray had cached an accessor instance on a Dataset. Renamed each to a unique class name. 2. The Optuna trial callback in tuning.py was a partial that captured train[0] as an unused kwarg. The partial was stored on the model via set_params(callback=...) and pickled along with it, dragging the whole training dataset (including any cached accessors) into the saved state. Dropped the unused capture and the dead code that referenced it. Also added TopographyModel.__getstate__ to null out `callback` before pickling, so any future closures captured by training-loop hooks cannot pin training data into a saved model.
Source of model instability has been identified, so the per-iteration finite checks, log-space max-subtract, and clip(-30, 30) added in f17c3c7 are no longer needed and add overhead in the offsets path.
`v / 1` was a debugging hack accidentally committed in d78a4f6; restored to `v / extent` so bars normalize to the global max and fit within the [0, 1.1] ylim.
- 31 tests covering inference, gtensor ops, save/load, plotting, and CLI - fixtures hosted on sigscape/MuTopia 'test-fixtures' release; conftest downloads on first use and caches under tests/fixtures/ - tests/build_chr22_fixture.py regenerates fixtures from tutorial data - tests.yml workflow runs on push and PR - publish.yml gated on tests passing before PyPI upload - joblib added as explicit dependency (was transitive via sklearn)
mutopia.__version__ is now read from package metadata rather than hard-coded in __init__.py. Tagging vX.Y.Z is the only step needed to release; setuptools_scm derives the wheel/package version from the tag at build time. Between tags, dev installs report e.g. 1.0.8.dev3+gabc1234 automatically. mutopia/_version.py is generated at build, gitignored.
Five new tests in tests/test_training.py exercise full-batch and SVI
training, save/reload after fit, and init_components with COSMIC names.
They are skipped by default; run with:
pytest tests/test_training.py --runslow
Conftest grows a --runslow option and a 'slow' marker. CI keeps the
fast tier (no training) so PR checks stay under 30s.
Note: test_batch_training_test_scores_nondecreasing currently fails on
the chr22-train / chr1-test fixture (test score degrades monotonically
across epochs). Likely distributional mismatch between fixture chroms
or overfitting at k=2; worth investigating before relying on this as
a regression gate.
… Zenodo
Previous training tests used the chr22 fixture, which is too small and
distributionally dissimilar from the chr1 test set to give meaningful
convergence signal. Slow-tier tests now download Liver-HCC.nc (~540 MB)
from Zenodo record 18803136, cache to tests/fixtures/zenodo/, and split
on chr1 for held-out evaluation.
Run with:
pytest tests/test_training.py --runslow
Default fast tier and CI are unchanged.
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
1 participant
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.
No description provided.