reshape_op

→TensorImplPtr

int reshape_op(const int & a, const int & new_shape)

source edit

Python wrappers

pylucid.ops.reshapefunction

Reinterpret a tensor under a new shape (no data copy).

The total element count must be preserved. Exactly one entry of new_shape may be -1, in which case that dimension is inferred from the remaining sizes. The input must be contiguous; pass non-contiguous inputs through contiguous_op first.

Shape

$\mathrm{numel}(a) = \prod_d \text{new\_shape}[d]$ , after resolving any -1 entry.

See Also

flatten_op, squeeze_op, contiguous_op.

Parameters

aTensorImplPtr

Input tensor. Must be C-contiguous.

new_shapevector<int64_t>

Target shape. At most one entry may be -1. All other entries must be non-negative. Since Shape == vector<int64_t>, this overload also accepts Shape directly.

Returns

TensorImplPtr

A view tensor with shape new_shape (with -1 resolved) sharing storage with a.

Raises

ShapeMismatch

If element counts do not match, multiple -1 entries are given, or the input is non-contiguous.

Examples

reshape_op(t_2x3, {6}) flattens a 2x3 to a 1-D length-6 view.
reshape_op(t_2x6, {-1, 3}) infers the first dim as 4.