dstack

→Tensor

dstack(tensors: Sequence[Tensor])

source

Stack tensors along the third (depth) axis.

Reshapes lower-rank inputs to give them a depth dimension, then concatenates along axis 2. 0-D scalars become (1, 1, 1), 1-D vectors of length N become (1, N, 1), and 2-D matrices of shape (H, W) become (H, W, 1).

Parameters

tensorsSequence[Tensor]

Tensors to stack. 0-, 1-, and 2-D inputs are auto-promoted to 3-D as described above; 3-D inputs pass through unchanged.

Returns

Tensor

3-D tensor formed by concatenating the (possibly promoted) inputs along axis 2.

Notes

The promotion rules ensure the result is always at least 3-D, which is useful for stacking colour channels of images: dstack([R, G, B]) produces an (H, W, 3) tensor from three (H, W) planes.

Examples

>>> import lucid
>>> a = lucid.tensor([[1., 2.], [3., 4.]])
>>> b = lucid.tensor([[5., 6.], [7., 8.]])
>>> lucid.dstack([a, b]).shape
(2, 2, 2)