Drop unused cg_lmax5.json data file#43
Closed
johmathe wants to merge 170 commits into
Closed
Conversation
…mmits Add pre commit and actions
Update general API and add cg coefficients
…ff, not the max value that l can take. Hence, l takes values in range(0, lmax). Need to take this into account when building l1 and l2 with the bsp.
Fix indices in the bsp
…e/fix Revert "Fix indices in the bsp"
Proposes a clean, mathematically faithful API covering:
- Group{on}Domain naming convention (CnonZn, DnonR2, SO3onS2)
- nn.Module-only interface with raw signal inputs
- float32 throughout
- Selective bispectrum as default
- Public API surface and file structure
- Open questions for team discussion
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Explicitly documents that the shift to raw signal inputs breaks SO3onS2 v0.1.0 interface, explains implications, and justifies why the tradeoff is worthwhile (consistency, encapsulation, correctness, testability). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Replaces inline Unicode math with proper $...$ and $$...$$ LaTeX syntax throughout the design document.
- SO(3) bispectrum formula: remove \hat accent on \mathcal{F} (caused
rendering issues with accent stacking on calligraphic letters); also
use proper indexed CG notation C_{l1,l2}^{(l)} matching the paper
- Q3 heading: fix \text{onS}^2 -> \text{ on }S^2 so the superscript
attaches to S not to the whole text box
- Migration table: remove incorrect escnn reference (we do not depend
on escnn; CG data is bundled as cg_lmax5.json)
- Scorecard table covering all group/domain combinations: Cn, Dn, O (selective + inversion proven), SO(3)/S², disk, ball, compact groups, homogeneous spaces (full bispectrum only, open problems) - Four mathematical TODOs for future selective versions: M1: SO(3) on S² (infinite irreps, BFS doesn't terminate) M2: SO(2) on disk (commutative structure, likely tractable) M3: SO(3) on ball (same challenge as S² + radial dim) M4: Inversion for continuous groups (truncation + stability)
- Replace $C_{l_1,l_2}^{(l)}$ inline (multi-underscore) with simplified prose
- Replace $(\rho_1,\rho_2)$ with $(\rho,\sigma)$ to avoid dual-underscore italic misparse
- Replace $O(|G|^2)$ with $O(\lvert G\rvert^2)$ using proper LaTeX delimiters
- Replace $\|G\|$ in table cell with $\lvert G \rvert$ to avoid table pipe conflict
- Replace $\mathbb{R}_{+}^{*}$ (underscore after brace) with $\mathbb{R}^+$ in table cells and headings
- Rewrite TODO-M1 paragraph: remove multi-subscript inline math, state O(L^3) cleanly
- Move Fourier-Bessel integral from inline to display math in TODO-M2
- Fix TODO-M3 heading notation to match TODO-M2
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…struction Add spherical reconstruction + update experiments methodology
# Conflicts: # paper/paper.pdf # paper/paper.tex
Captures completed sMNIST, Organ3D, and PCam runs from the H100 parallel sweep, including the new cost metadata (wall_time_s, epochs_run, peak GPU memory, gpu_name, git_sha, argv). Made-with: Cursor
…iments Experiment results
- Eliminate Python for-loop in DnonDn.forward by batching all CG-transformed bispectrum computations: precomputed Fplus gather indices, single einsum for kron products, batched matmul chain. Speedup: 2.41ms -> 0.55ms (4.4x). - Add test_coverage_gaps.py with 39 edge-case tests covering _bessel, _cg, dn_on_dn, so3_on_s2, octa_on_octa, rotation, and torus_on_torus modules. - Overall test coverage: 91% -> 96%. Co-authored-by: Cursor <cursoragent@cursor.com>
…dnondn-and-coverage Optimize DnonDn forward pass and increase test coverage to 96%
- Use importlib.metadata for __version__ - Pin torch>=2.10 (real ABI floor for torch_harmonics 0.9.0) - Pin torch-harmonics>=0.9 - Add LICENSE file (MIT) - Restructure extras: drop [plotting], add [experiments] with matplotlib/scipy/torchvision/h5py/medmnist
Add publish/RELEASE
Release v0.3.5
- Restructure paper/experiments/ -> experiments/
- Apply experiments code edits (organ3d, pcam, spherical_mnist, spherical_mnist_reconstruction)
- Enrich README with Supported Groups, API, and Benchmarks sections
- Delete docs/, DESIGN.md, benchmarks/figures/, benchmarks/inversion_results/,
benchmarks/paper_benchmarks.{tex,pdf}, .cursorrules, paper/ working files
- Exclude experiments/ from pre-commit hooks (research scripts, not library code)
Preserved on main: LICENSE, RELEASE.md, .github/workflows/publish.yml,
pyproject.toml authors/URLs/version (0.3.5), .github/workflows/tests.yml
codecov slug, README CI badges and clone URL.
Co-authored-by: Cursor <cursoragent@cursor.com>
…ips2026anonymous Merge improvements from neurips2026anonymous (sans anonymization)
The JSON of precomputed Clebsch-Gordan coefficients was loaded into
_DATA_PATH in _cg.py but never read. CG values are computed analytically
via the Wigner 3j symbol (Racah formula) at runtime, with on-disk caching
under ~/.cache/bispectrum/.
- Delete src/bispectrum/data/{cg_lmax5.json,__init__.py}
- Drop _DATA_PATH definition in _cg.py
- Update _cg.py module docstring (no precomputed-load support)
- Drop *.json from hatch include glob in pyproject.toml
Wheel size: ~1.55 MB -> ~174 kB.
Co-authored-by: Cursor <cursoragent@cursor.com>
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.
Summary
`src/bispectrum/data/cg_lmax5.json` (1.3 MB) was bundled in the wheel but never read at runtime. `_DATA_PATH` was defined in `_cg.py` and orphaned. Clebsch-Gordan coefficients are computed analytically via the Wigner 3j symbol (Racah formula) and cached to disk under `~/.cache/bispectrum/`.
Wheel size drops from ~1.55 MB to ~174 kB.
Test plan
Made with Cursor