Table of Contents

Fundamentals and usage

These pages clarify the concepts, design decisions, and behavioral rules behind NumSharp - the .NET port of NumPy that targets 1-to-1 API and behavioral compatibility with NumPy 2.x. They are the NumSharp counterparts of NumPy's own NumPy fundamentals guide: one article per NumPy article, rewritten for C# and honest about where a .NET runtime forces NumSharp to diverge.

If you have written NumPy before, these pages are the fastest way to map that knowledge onto NDArray. If you have not, they are the place to build a mental model before reaching for the API reference.

The articles

Article What it covers NumPy original
Array creation The six ways to make an NDArray - from .NET sequences, intrinsic functions, joins, disk, raw bytes, and random basics.creation
Indexing on NDArray Basic indexing (views), advanced indexing (copies), boolean masks, the flat iterator, and assignment basics.indexing
I/O with NumSharp The byte-exact .npy/.npz stack, text I/O (loadtxt/savetxt), and raw-binary I/O basics.io
Data types Numeric and text descriptors, NumPy mappings, promotion, and the .NET-specific Decimal and Char types basics.types
Broadcasting How arrays of different shapes combine in arithmetic - the shape rules and the stride-0 mechanism basics.broadcasting
Copies and views Which operations share memory and which duplicate it, and how to tell basics.copies
Strings, bytes and chararrays Fixed-width S/U, variable-width StringDType, missing values, text computation, numpy.char and NDCharArray; how they differ from C# Char basics.strings
Structured arrays Why NumSharp does not implement record/structured dtypes, and the .NET-idiomatic alternatives basics.rec
Universal functions (ufunc) basics Elementwise ops, out=/where=/dtype=, broadcasting, casting, and reductions basics.ufuncs

Data types and Broadcasting already have dedicated, in-depth NumSharp guides (Dtypes, Broadcasting); they are the NumSharp conversions of NumPy's data-types and broadcasting articles, so they are linked here rather than duplicated.

What "convert each NumPy article to a NumSharp one" means

NumSharp's goal is that ported NumPy code behaves identically. So each article follows its NumPy source closely - same concepts, same section order, the same worked examples - but:

  • Code is C#, not Python. np.zeros((2, 3)), arr["1:3, :2"], arr.astype(np.float64). Python slice literals (a[1:5:2]) become strings in C# (a["1:5:2"]) because C# has no slice syntax - see Indexing.
  • NumSharp's dtypes follow NumPy's model. Arrays can hold numbers, Unicode strings, byte strings or variable-width StringDType text; Decimal and Char add .NET-specific choices. Object, void and structured arrays, native complex64 storage and datetime/timedelta storage are not represented. Each article explains the relevant rules and alternatives.
  • Divergences are documented, never silent. Where NumSharp cannot match NumPy exactly (e.g. a C# int[] is a strong array that wraps on downcast, where a Python list is weak and raises), the article flags it. Behavioral differences are gated in the test suite under the Misaligned category.

The fundamentals articles are conceptual. When you want the exhaustive how-to, the neighbouring guides go deeper:

Edit this page