[sfc0324]

In one sentence: vesuvius.neural_tracing.fiber_trace can be imported from a normal vesuvius install without the monorepo lasagna package on PYTHONPATH.

One real example: Starting with a fresh uv sync --extra models of vesuvius at a4979b9, I ran tests/neural_tracing/test_fiber_trace.py, and it produced 52 passing tests instead of a collection-time ModuleNotFoundError: No module named 'lasagna'.

Before: fiber_trace/geometry.py imports _decode_normals from lasagna.omezarr_pyramid, falling back to a bare omezarr_pyramid module. lasagna is not a declared dependency of vesuvius, so importing vesuvius.neural_tracing.fiber_trace (and collecting its test file) fails unless the user manually puts villa/lasagna on PYTHONPATH. Only this 4-line decoder is used from lasagna here.

After this PR: lasagna's implementation is still preferred when importable; otherwise an identical local _decode_normals_u8 is used. Behaviour with lasagna installed is unchanged.

Proof: Same environment, base a4979b9 vs this branch.

Before (a4979b9):

src/vesuvius/neural_tracing/fiber_trace/geometry.py:9: in <module>
E   ModuleNotFoundError: No module named 'omezarr_pyramid'
ERROR tests/neural_tracing/test_fiber_trace.py
!!!!!!!!!!!!!!!!!!!! Interrupted: 1 error during collection !!!!!!!!!!!!!!!!!!!!
1 error in 1.62s

After (this PR; test_fiber_trace.py + new test_fiber_trace_decode_normals.py):

55 passed in 5.61s

The 3 new tests: known values (128→0, 255→1, 1→−1, float32), byte-for-byte parity with lasagna/omezarr_pyramid.py::_decode_normals on random uint8 input (loaded from the monorepo file; skips if absent), and import with lasagna/omezarr_pyramid blocked in sys.modules.

Why / where this is useful: People who pip/uv install vesuvius to train or run the fiber tracer, and CI/test runs of the vesuvius package on its own, no longer need the undocumented PYTHONPATH=lasagna step for this module.

• I personally verified that the example and proof above were produced by this PR on the stated data.

Details

• Tested commit: base a4979b9; this branch = one commit on top.
• Environment: Linux x86_64, CPU only (no GPU), no C++ compiler on PATH (which g++ c++ finds nothing), Python 3.14.8, torch 2.12.1+cu130, uv 0.13.0. Install: cd vesuvius && uv sync --extra models --no-install-package volume-cartographer (volume-cartographer skipped because it needs a C++ toolchain to build).
• Command (from vesuvius/): uv run --with pytest python -m pytest -q -p no:cacheprovider tests/neural_tracing/test_fiber_trace.py tests/neural_tracing/test_fiber_trace_decode_normals.py
• Changes: src/vesuvius/neural_tracing/fiber_trace/geometry.py (local decoder + third import fallback); new tests/neural_tracing/test_fiber_trace_decode_normals.py.
• Limitations / not changed: fiber_trace_2d and fiber_trace_3d use much more of lasagna (tiled_predict3d, live_omezarr_cache, ...) and still require it; their test files still error at collection without lasagna. Keeping a copy of the decoder means the two must stay in sync — the parity test guards that.

[vercel]

@sfc0324 is attempting to deploy a commit to the scroll Team on Vercel.

A member of the Team first needs to authorize it.

[hendrikschilling]

Looks mostly fine, but this PR duplicates the lasagna normal decoder in vesuvius. To avoid duplication please move it into vesuvius instead and make lasagna code use this shared decoder instead.

[sfc0324]

Thanks for the review! I moved the normal decoder into vesuvius (vesuvius.neural_tracing.normals.decode_normals). fiber_trace now uses it directly, and lasagna.omezarr_pyramid re-exports it in place of its local copy, so existing imports keep working. Added tests for the shared decoder, and the vesuvius and lasagna tests pass.