v0.21.1: TT _dim_order cluster + open issues + perf#23
Merged
Conversation
Raise ValueError with reorder() hint when self._dim_order != other._dim_order instead of silently computing a meaningless Frobenius product. Mirrors the v0.20.1 binary algebra behavior on _check_compatible_tt mismatches. Closes ultrareview Critical finding.
Replace exact == / != on .domain and .n_nodes with np.allclose / np.array_equal on coerced arrays. Tolerates tuple-of-tuples vs list-of-lists syntax that silently arises in scalar-mul (which normalizes via [list(b) for b in domain]). Closes issue #22.
Permute columns by inverse _dim_order before returning so that eval(get_evaluation_points()[i]) round-trips for any TT regardless of storage permutation. Matches Approximation/Spline/Slider behavior. Closes ultrareview Important finding. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add _user_frame_domain() private helper that returns self.domain permuted into user-frame order. The three v0.21 methods now pass this user-frame view to _calculus._validate_calculus_args instead of the raw storage-frame self.domain. Pre-fix: under non-identity _dim_order with non-uniform per-dim domains, validation referenced the wrong domain. Tests pass before the fix only because uniform domains masked the difference. Closes ultrareview Critical finding. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add optional dim_labels= parameter to _normalize_bounds. ChebyshevTT.integrate passes user-frame dims_sorted as dim_labels so out-of-domain bounds errors reference the dim the user passed rather than the storage-frame index. Helper stays user-frame-naive otherwise. Closes #20. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ssue #19) Extract _eval_storage_frame(point_storage, deriv_storage) private helper that takes already-permuted coords. eval() permutes user→storage once at entry and calls _eval_storage_frame; eval_multi() permutes once and calls the helper N times for each derivative_order. Drops the try/finally self._dim_order mutation entirely. Eliminates the race rather than papering over it with a lock. The FD machinery (_fd_single_dim, _fd_cross_deriv, _fd_nested) was updated to call _eval_storage_frame instead of self.eval, since the points it operates on are already in storage frame; routing through self.eval would re-permute and break correctness once the saved-state trick is removed. Closes #19.
Replace the Python list comprehension over candidates with a single vectorized barycentric evaluation (BLAS GEMV pattern). _optimize_1d is the canonical 1-D optimizer used by all four classes' minimize/maximize methods; v0.21 routed Slider/TT min/max through it, making the perf hoist worthwhile. Bit-identical results — non-regression tests pass. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…atch loop The differentiation matrix (D_T) matmul is point-independent — applying it once to self.tensor_values before the per-point loop, then doing only the barycentric reduction inside the loop, produces identical results with substantial speedup for batch derivative evaluations. Each D_T[d] is applied along axis d of the full tensor (via moveaxis to last, matmul, moveaxis back) to correctly match the per-dim interleaved semantics of vectorized_eval. ChebyshevSpline.eval_batch (which delegates per-piece) inherits the speedup.
Demonstrates each fix in the TT _dim_order cluster with a non-uniform domain and explicit reorder() — the configuration that v0.20.1/v0.21.0 test coverage masked. Native TT sobol_indices deferred to v0.22; not included in this demo.
Add non-uniform-domain validation note to user-guide. Bump version to 0.21.1. CHANGELOG entry covering all 6 fixes + 2 perf wins, with sobol_indices deferred to v0.22. CLAUDE.md updated with new test count and v0.21.1 architecture line.
|
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
… sites Final cumulative review (M1) noted that ChebyshevTT.inner_product and the binary-algebra _check_compatible_tt path also use exact == on self.domain, paralleling the issue #22 fix in _algebra._check_compatible. Apply the same np.allclose pattern to both TT sites for consistency with the v0.21.1 fix. (M2) Add compare_v0211_dim_cluster.py to the CLAUDE.md benchmark roster.
Compute first-order + total-order Sobol indices by contracting through the TT coefficient cores in O(d * n * r^2) time. Cross-validated against ChebyshevApproximation.sobol_indices() on dense reconstructions. Closes the v0.21.1 sobol_indices parity item (originally deferred but attempted on user request).
Appends test_non_uniform_domain_after_reorder to TestTTSobolParity, covering the exact configuration that masked v0.21.0 _dim_order bugs: non-uniform per-dim domains [(-1,1), (-2,2), (-3,3)] combined with a non-identity _dim_order (reorder([2,0,1])). Verifies TT sobol keys stay in user-frame after reorder, matching ChebyshevApproximation reference. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…Issue 2) Adds private _apply_derivative_passes(tensor, derivative_order) that applies differentiation-matrix passes along each axis of the full coefficient tensor using np.moveaxis. Used by vectorized_eval_batch to replace the duplicated ~10-line derivative-hoist block, making the batch path self-documenting. vectorized_eval inlines its own per-axis logic (reduces axes as it goes, so the full-tensor moveaxis path doesn't apply there) and is left unchanged to avoid unneeded churn. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…r sobol (Issue 3) Replaces d independent full self-inner-product loops for total-order computation with precomputed L[k] (left partial, dims 0..k-1) and R[k] (right partial, dims k..d-1) matrices. For each dim j, combines L[j], the alpha_j=0 core slice, and R[j+1] via a single einsum instead of re-running all d cores. Reduces from O(d^2*n*r^2) to O(d*n*r^2). Existing TestSobolFromTTCores and TestTTSobolParity tests continue to pass at 1e-9 (2-D) and 1e-7 (3-D) tolerances. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.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
Closes the v0.20+v0.20.1
_dim_ordercluster onChebyshevTTsurfaced by ultrareview, plus three open follow-up issues, plus two perf hoists.Critical/Important fixes (TT
_dim_ordercluster)roots/minimize/maximizevalidatefixedagainst user-frame domain (was storage-frame; under non-uniform domains + non-identity_dim_ordercould raise misleading errors or accept invalid inputs)inner_productraisesValueErroron mismatched_dim_order(was silently miscomputing Frobenius product)get_evaluation_pointsreturns columns in user-frame order (was storage order, breakingeval(get_evaluation_points()[i])round-trip)Open issues closed
#19—eval_multithread-safety: structural fix via_eval_storage_frameprivate helper (also reroutes_fd_*helpers)#20—integrateerror message: now references user-frame dim index#22—_algebra._check_compatible: now usesnp.allclose(handles tuple-vs-list mixed domain syntax)Performance
_calculus._optimize_1dvectorized — used by all four classes'minimize/maximizevectorized_eval_batchderivative-matrix hoist (with moveaxis correction for non-trailing axes) — speedup for batch derivative evaluationsDeferred
ChebyshevTT.sobol_indicesdeferred to v0.22 (native TT contraction implementation exceeded v0.21.1 time budget)Tests
~25 new tests across
test_calculus_completion.py,test_v0201_dim_threading.py,test_tensor_train.py,test_calculus.py,test_algebra.py,test_barycentric.py. Total project: ~1133 passing.Test plan
uv run python compare_v0211_dim_cluster.pyshows all[OK ]mkdocs build --strictcleanuv buildproduces 0.21.1 wheel + sdist🤖 Generated with Claude Code