kron_op

→TensorImplPtr

int kron_op(const int & a, const int & b)

source edit

Python wrappers

pylucid.ops.kronfunction

Kronecker (block tensor) product over matching ranks.

Composite over reshape_op + mul_op + reshape_op. Interleaves size-1 placeholders so the multiply broadcasts to a $2 \cdot \mathrm{ndim}$ -rank block tensor; the trailing reshape collapses each (sa[i], sb[i]) pair into a single sa[i] * sb[i] axis. Gradient flows through ReshapeBackward and MulBackward.

Math

For rank-2 operands $A \in \mathbb{R}^{m \times n}$ and $B \in \mathbb{R}^{p \times q}$ :

(A \otimes B)_{ip + r,\, jq + s} = A_{ij} \, B_{rs}

generalised elementwise to arbitrary equal ranks.

See Also

mul_op, reshape_op.

Parameters

a, bTensorImplPtr

Operands sharing the same rank ndim. Shapes need not match dimension-by-dimension.

Returns

TensorImplPtr

Tensor of shape (sa[0] * sb[0], ..., sa[-1] * sb[-1]).

Raises

Failure

If either input is null or the ranks differ.

Examples

kron of two (2, 3) matrices yields a (4, 9) matrix.