Overview
NumSharp ships np.unique (with the full return_index/return_inverse/return_counts/axis overload) but none of NumPy's set-operation routines built on top of it. This tracks adding the 9 missing numpy set functions -- every one is a thin composition over primitives NumSharp already has, so there are no new kernels to write.
Problem
NumPy's Set routines category (numpy/lib/_arraysetops_impl.py) exposes 10 functions; NumSharp implements 1 (unique). The membership / intersection / union / difference operations -- bread-and-butter for label handling, dedup, and data wrangling -- are all absent, along with the NumPy 2.x Array-API unique_* variants.
Verified codebase state
- Implemented:
np.unique -- including unique(return_index, return_inverse, return_counts, axis, equal_nan) (Manipulation/np.unique.cs).
- Missing (all 9 below): verified absent from the
np.* surface (reflection over the compiled np) and as NDArray methods / internal helpers (source grep across src/NumSharp.Core).
Proposal
Add the 9 functions under the main np.* namespace (NumPy keeps them there; several are re-exported in np.lib). All are pure composition -- the unique_* family is one-liners because unique's return flags already exist; the binary ops compose unique/sort/searchsorted.
Task checklist (ordered by real-world usage)
Evidence -- GitHub usage ranking
| # |
Function |
GitHub hits |
What it does |
Implementable via |
| 1 |
np.isin |
33,936 |
element-wise membership test element in test_elements (broadcasts over element) |
unique+searchsorted |
| 2 |
np.intersect1d |
22,168 |
sorted unique elements common to both arrays |
unique+searchsorted |
| 3 |
np.setdiff1d |
20,912 |
sorted unique elements of ar1 not in ar2 |
built on isin |
| 4 |
np.union1d |
15,824 |
sorted unique union of two arrays |
unique(concatenate) |
| 5 |
np.unique_values |
4,440 |
Array-API: the unique values only |
unique wrapper |
| 6 |
np.unique_inverse |
4,056 |
Array-API: unique values + inverse-reconstruction indices |
unique(return_inverse) |
| 7 |
np.unique_counts |
3,824 |
Array-API: unique values + counts |
unique(return_counts) |
| 8 |
np.setxor1d |
2,496 |
sorted unique symmetric difference |
unique+count |
| 9 |
np.unique_all |
2,068 |
Array-API: unique values + indices + inverse + counts |
unique(return_index,inverse,counts) |
Combined usage ~ 109,724 hits. isin alone (~34K) is among the most-used missing functions across the whole set+sort+indexing sweep.
Usage = GitHub REST code-search total_count in each function's canonical call form (np.<fn>(, or np.<fn>[ for the index-expression objects). It is an approximate estimate (noisy plus/minus ~10% on large sets) and counts only the dominant np. alias. Measured 2026-07-19; treat as relative demand, not exact.
Scope / Non-goals
- Composition only -- no new IL kernels; reuse
unique/sort/searchsorted.
in1d (deprecated pre-2.0 alias of isin) is not in scope.- 1-D semantics per NumPy (inputs raveled unless documented otherwise);
axis= only where NumPy supports it.
Related issues
Overview
NumSharp ships
np.unique(with the fullreturn_index/return_inverse/return_counts/axisoverload) but none of NumPy's set-operation routines built on top of it. This tracks adding the 9 missingnumpyset functions -- every one is a thin composition over primitives NumSharp already has, so there are no new kernels to write.Problem
NumPy's Set routines category (
numpy/lib/_arraysetops_impl.py) exposes 10 functions; NumSharp implements 1 (unique). The membership / intersection / union / difference operations -- bread-and-butter for label handling, dedup, and data wrangling -- are all absent, along with the NumPy 2.x Array-APIunique_*variants.Verified codebase state
np.unique-- includingunique(return_index, return_inverse, return_counts, axis, equal_nan)(Manipulation/np.unique.cs).np.*surface (reflection over the compilednp) and asNDArraymethods / internal helpers (source grep acrosssrc/NumSharp.Core).Proposal
Add the 9 functions under the main
np.*namespace (NumPy keeps them there; several are re-exported innp.lib). All are pure composition -- theunique_*family is one-liners becauseunique's return flags already exist; the binary ops composeunique/sort/searchsorted.Task checklist (ordered by real-world usage)
isin(33,936) -- element-wise membership testelement in test_elements(broadcasts overelement)intersect1d(22,168) -- sorted unique elements common to both arrayssetdiff1d(20,912) -- sorted unique elements ofar1not inar2union1d(15,824) -- sorted unique union of two arraysunique_values(4,440) -- Array-API: the unique values onlyunique_inverse(4,056) -- Array-API: unique values + inverse-reconstruction indicesunique_counts(3,824) -- Array-API: unique values + countssetxor1d(2,496) -- sorted unique symmetric differenceunique_all(2,068) -- Array-API: unique values + indices + inverse + countsEvidence -- GitHub usage ranking
np.isinelement in test_elements(broadcasts overelement)unique+searchsortednp.intersect1dunique+searchsortednp.setdiff1dar1not inar2isinnp.union1dunique(concatenate)np.unique_valuesuniquewrappernp.unique_inverseunique(return_inverse)np.unique_countsunique(return_counts)np.setxor1dunique+countnp.unique_allunique(return_index,inverse,counts)Combined usage ~ 109,724 hits.
isinalone (~34K) is among the most-used missing functions across the whole set+sort+indexing sweep.Usage = GitHub REST code-search
total_countin each function's canonical call form (np.<fn>(, ornp.<fn>[for the index-expression objects). It is an approximate estimate (noisy plus/minus ~10% on large sets) and counts only the dominantnp.alias. Measured 2026-07-19; treat as relative demand, not exact.Scope / Non-goals
unique/sort/searchsorted.in1d(deprecated pre-2.0 alias ofisin) is not in scope.axis=only where NumPy supports it.Related issues