meshgrid_op

→vector<TensorImplPtr>

int meshgrid_op(const int & xs, bool indexing_xy)

source edit

Python wrappers

pylucid.ops.meshgridfunction

Build N broadcast coordinate grids from N 1-D input tensors.

Given inputs $x_0, x_1, \ldots, x_{N-1}$ with lengths $L_0, L_1, \ldots, L_{N-1}$ , produces N output tensors all of the same rank-N shape; the $i$ -th output broadcasts $x_i$ across every axis except its "carry" axis, where it varies. This is the multi-dimensional analogue of NumPy's np.meshgrid.

Math

For ij indexing: $\mathrm{out}_i[k_0, k_1, \ldots, k_{N-1}] = x_i[k_i]$

See Also

Concat : Stack the resulting grids along a new leading axis.

Parameters

xsvector<TensorImplPtr>

One or more 1-D input tensors. All must share the same dtype and device.

indexing_xybool

Axis convention for the first two inputs: * false — "ij" (matrix) indexing. Output i varies along axis i; output shape is (L_0, L_1, ..., L_{N-1}). * true — "xy" (Cartesian) indexing. Output 0 varies along axis 1 (x) and output 1 varies along axis 0 (y); outputs i >= 2 follow the ij convention. Output shape is (L_1, L_0, L_2, ..., L_{N-1}).

Returns

vector<TensorImplPtr>

N coordinate grids, one per input, each of rank N.

Notes

Backward sums each output gradient over every axis except the carry axis to recover the gradient w.r.t. the corresponding 1-D input. All inputs must be 1-D; mismatched ranks raise.

Examples

>>> X, Y = meshgrid([x, y], indexing_xy=true)  // 'xy' convention
>>> X.shape == (len(y), len(x))
True