feat(python): add GUNTAM Python-callable seeding example#5551
Conversation
Corentin-Allaire
left a comment
Corentin-Allaire
left a comment
There was a problem hiding this comment.
Here are a few comments on the PR. One thionk I am not 100% sure is if we want to push Examples/Scripts/Python/onnx_seeding.py as is to the repo ?
EliMe5
force-pushed
the
custom-python-seeding-implementation
branch
2 times, most recently
from
f6426a7 to
54aa262
Compare
|
drafting for now until we have a conclusion on #5571 which takes load off the gitlab ci |
@benjaminhuth I could try, but the code requires some of the guardrails (particularly the missing |
- Add `SeedingAlgorithm.PythonCallable` configuration path. - Introduce `customSeeder` and `customSeederConfig` parameters. - Provide general interface for external/ML seeding functions.
- r_max, z_max, score_threshold, and providers are now config parameters with the previous hardcoded values as defaults (500 mm, 1000 mm, 0.35, CPUExecutionProvider). - Remove z-degeneracy filter (_filter_degenerate_triplets / _Z_EPS): no longer needed after binding lifetime fixes removed the defensive crash- avoidance path. - Move duplicate-index check into _filter_valid_seeds (sanity check, not a physics cut). - Pre-allocate all seed proxies in one pass before filling to avoid repeated container growth inside the fill loop. - GPU support: pass providers=["CUDAExecutionProvider", "CPUExecutionProvider"] to use GPU with CPU fallback.
…exing Replace the per-seed Python loop over orig_idx with a single NumPy fancy indexing operation (orig_idx[seeds_v], shape N×3), then call .tolist() per row inside the fill loop.
SpacePointIndex2 = std::uint32_t; using int64 caused an implicit sign/range coercion on every assignSpacePointIndices call.
Adds a transverse-separation degeneracy check to _filter_valid_seeds to drop seeds where two spacepoints are numerically coincident in the xy plane — these produce degenerate track parameters that crash the CKF at high event counts. Also renames the internal 'hits' variable to 'filtered_sp_xyz' to avoid confusion with Acts::SimHit (truth-level pre-digitization data).
EliMe5
force-pushed
the
custom-python-seeding-implementation
branch
from
54aa262 to
aa2dcac
Compare
|
benjaminhuth
left a comment
benjaminhuth
left a comment
There was a problem hiding this comment.
Looks good! Only minor comment (maybe also a bit matter of taste): I think right now it onnx_seeding.py suggests some generality which it does not have, its really just running guntam. So maybe merging the two py files into one called guntam_transformer_seeder.py makes sense?
But not critical.
4 participants
feat(python): add GUNTAM Python-callable seeding example
Add a PythonCallable seeding mode that allows reconstruction seeding to be
provided by a Python function.
Expose the event-data bindings needed by Python seeders to construct seeds
from Python, including mutable space point proxies, mutable seed proxies,
seed-spacepoint container assignment, and tests for the ownership path.
Add a GUNTAM ONNX seeding example using the OpenDataDetector workflow. The
integration is implemented for GUNTAM, while the PythonCallable interface
and ONNX example structure are intended to support other externally provided
ONNX seeding models as well.
--- END COMMIT MESSAGE ---
This PR adds:
SeedingAlgorithm.PythonCallablesupport through the reconstruction helpersSpacePointContainer2/MutableSpacePointProxy2bindingsSeedContainer2/MutableSeedProxy2bindings, including seed creation and space-point container assignmentIndexSourceLinkconstructor andmakeSourceLinkbindings in the examples moduleLocal validation:
origin/mainpytest Python/Core/tests/test_event_data.pypassedpy_compilepassed on both example scriptsonnx_seeding.pywith a GUNTAM model for 2 particle-gun events, single thread; full chain completed successfully through Fatras, digitization, seeding, CKF, ambiguity resolution, and vertex fittingcc @benjaminhuth @Corentin-Allaire