- Generated by
1.9.8
|
Stratax 0.4.0rc1
|
The public package re-exports the compiled _core API. Numeric Vector, Matrix, and Tensor store double values. Boolean comparison results and int64 index results have separate classes with narrower interfaces.
| Type | Constructors |
|---|---|
| Shape | Shape(), Shape(other), Shape(size), Shape([d0, d1, ...]) |
| Vector | Vector(), Vector(other), Vector(size), Vector(size, value), Vector(shape), Vector(values) |
| Matrix | Matrix(), Matrix(other), Matrix(rows, cols), Matrix(rows, cols, value), Matrix(shape), Matrix(nested_rows) |
| Tensor | Tensor(), Tensor(other), Tensor(shape), Tensor(shape, value) |
Tensor accepts a Shape or an iterable of dimensions, not nested element data. Tensor([2, 3]) creates six zeros. Matrix rows must be rectangular. Dimensions must be nonnegative integers. Checked scalar construction and element assignment reject bool and nonfinite values; arithmetic results follow core floating-point rules and can be nonfinite.
Shape exposes rank, elements, empty, strides, dimension indexing, iteration, and equality. Shape(size) creates a rank-one shape. Shape.strides computes row-major element strides as another Shape. Shape.empty means rank zero. Shape.elements is zero for rank zero or any zero extent. Owning containers expose size, rank, empty, shape, strides, fill(value), tolist(), reshape(shape), and flatten(); Matrix adds rows and cols. Strides are in elements. Metadata are properties, not method calls.
Owning numeric len(arr) is the first extent, or zero for rank zero; iteration visits all elements in flat logical order. tolist returns nested lists matching the shape as far as a list can express it. A zero outer dimension becomes [] and cannot preserve unseen trailing dimensions in the list representation. Default Vector/Tensor have shape [0]; default Matrix has shape [0, 0]. Explicit Tensor(Shape()) is empty rank zero, not a scalar.
Vector integer access is flat. Matrix accepts (row, col) for read/write and an integer for flat reads; Matrix assignment requires a row/column tuple. Tensor supports flat integers and full-rank integer tuples. Negative indices count from the end. Rank-zero arrays have no addressable elements: reading or writing with an empty tuple raises stratax.IndexError. A tuple whose length differs from the array rank raises stratax.RankError.
Positive-step slices return shared ArrayView. Vector takes one slice, Matrix supports row slicing or a pair of indices/slices, and Tensor takes one entry per axis (a bare slice works only at rank one). Integer entries mixed with slices retain length-one axes. Slice bounds clamp to the extents. Negative steps raise stratax.IndexError; Python rejects a zero slice step with built-in ValueError. Bulk slice assignment and advanced indexing are not bound.
Views keep their owner alive and expose size, rank/ndim, empty, shape, strides, tolist(), and checked integer/tuple reads and writes. len(view) counts all selected elements. Views have no public constructor, further slicing, arithmetic, reshape, conversion, or reduction overloads. Use an explicit owning copy when an operation needs a numeric container.
+, -, *, /, unary +/- and compound arithmetic accept owning numeric containers; binary forms also accept scalars in either order and mixed container families. Array-array operations broadcast trailing dimensions. Matching families retain their family; mixed families return Tensor. Compound operations preserve the left object's shape and dtype and reject expansion with BroadcastError. Compound division validates all used divisors before writing. Scalar division by zero raises ZeroDivisionError even for an empty input. An empty rank-zero operand cannot supply values to a nonempty result.
==, !=, <, <=, >, and >= return BoolVector, BoolMatrix, or BoolTensor masks with the broadcasted shape. Named functions equal, not_equal, less, less_equal, greater, and greater_equal expose the same operations. At least one operand must be a supported numeric array. These are element-wise comparisons, not whole-array equality or approximate equality.
Boolean result classes expose metadata, fill, tolist, repr, flat iteration, and len (total element count). They do not expose public constructors or indexed access. Their truth value is not a reduction of the Boolean elements; use Python all/any over iteration for a whole-mask check.
| Function | Output selection |
|---|---|
zeros(shape), ones(shape), full(shape, value) | Tensor; shape must be a Shape |
zeros(size), ones(size), full(size, value) | Vector |
zeros(rows, cols), ones(rows, cols), full(rows, cols, value) | Matrix |
identity(size) | Square Matrix |
to_vector(arr) | Vector-compatible shape: rank one or exactly one extent different from one |
to_matrix(arr) | Matrix-compatible shape: rank two or exactly two extents different from one |
to_tensor(arr) | Tensor with unchanged shape |
arr.reshape(shape) | Tensor with equal element count |
arr.flatten() | Vector in flat logical order |
Creation, conversion, reshape, and flatten allocate independent storage. Conversion accepts owning numeric arrays. Only extent-one axes are removed; zero extents survive. Reshape does not infer -1 dimensions.
All nine reductions accept owning numeric arrays. Use (arr) globally or (arr, axis, keepdims=False) for an axis; axis=None and multi-axis sequences are not bound. Axes may be negative. keepdims retains extent one; reducing rank one without keepdims still returns shape [1].
| Functions | Global result | Axis result | Empty reduction slice |
|---|---|---|---|
| sum | float | Tensor | 0 |
| prod | float | Tensor | 1 |
| min, max | float | Tensor | stratax.IndexError |
| argmin, argmax | int | IndexTensor | stratax.IndexError |
| mean, var, std | float | Tensor | stratax.ZeroDivisionError |
An empty output domain produces an empty result without reducing any slices. Extrema use the first tied position. Global indices are flat; axis indices are positions along that axis. IndexTensor stores int64 and exposes metadata, fill, tolist, iteration, membership, len (first extent), and checked flat/tuple reads. It has no public constructor or indexed assignment. Variance/std are population statistics dividing by n, with no ddof or NaN-skipping option.
StrataxError derives from built-in RuntimeError. Exported subclasses are ShapeError, DimensionError, RankError, IndexError, TypeError, BroadcastError, ZeroDivisionError, AxisError, OverflowError, and ValueError. RankError handles wrong index/slice counts; AxisError handles invalid reduction axes; DimensionError includes shape arithmetic overflow.
These are distinct from Python's built-in exceptions of the same names. Binding conversion overflow and Shape.elements overflow explicitly raise built-in OverflowError. Shape.strides overflow raises stratax.DimensionError. Python slice parsing and pybind11 argument matching may raise built-in ValueError or TypeError. C++ allocation failures may become MemoryError. Catch the documented category at the relevant call, rather than assuming every error is StrataxError.
C++ bitwise operators, standard math wrappers, astype, and most additional dtypes are not exported by the Python package. See User Guide and Bindings Developer Docs for implementation boundaries.
1.9.8