SpaceToBatch for N-D tensors of type T.
tf.space_to_batch_nd(
input: Annotated[Any, TV_SpaceToBatchND_T],
block_shape: Annotated[Any, TV_SpaceToBatchND_Tblock_shape],
paddings: Annotated[Any, TV_SpaceToBatchND_Tpaddings],
name=None
) -> Annotated[Any, TV_SpaceToBatchND_T]
This operation divides "spatial" dimensions [1, ..., M]
of the input into a
grid of blocks of shape block_shape
, and interleaves these blocks with the
"batch" dimension (0) such that in the output, the spatial dimensions
[1, ..., M]
correspond to the position within the grid, and the batch
dimension combines both the position within a spatial block and the original
batch position. Prior to division into blocks, the spatial dimensions of the
input are optionally zero padded according to paddings
. See below for a
precise description.
This operation is equivalent to the following steps:
Zero-pad the start and end of dimensions
[1, ..., M]
of the input according topaddings
to producepadded
of shapepadded_shape
.Reshape
padded
toreshaped_padded
of shape:[batch] + [padded_shape[1] / block_shape[0], block_shape[0], ..., padded_shape[M] / block_shape[M-1], block_shape[M-1]] + remaining_shape
Permute dimensions of
reshaped_padded
to producepermuted_reshaped_padded
of shape:block_shape + [batch] + [padded_shape[1] / block_shape[0], ..., padded_shape[M] / block_shape[M-1]] + remaining_shape
Reshape
permuted_reshaped_padded
to flattenblock_shape
into the batch dimension, producing an output tensor of shape:[batch * prod(block_shape)] + [padded_shape[1] / block_shape[0], ..., padded_shape[M] / block_shape[M-1]] + remaining_shape
Some examples:
(1) For the following input of shape [1, 2, 2, 1]
, block_shape = [2, 2]
, and
paddings = [[0, 0], [0, 0]]
:
x = [[[[1], [2]], [[3], [4]]]]
The output tensor has shape [4, 1, 1, 1]
and value:
[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]
(2) For the following input of shape [1, 2, 2, 3]
, block_shape = [2, 2]
, and
paddings = [[0, 0], [0, 0]]
:
x = [[[[1, 2, 3], [4, 5, 6]],
[[7, 8, 9], [10, 11, 12]]]]
The output tensor has shape [4, 1, 1, 3]
and value:
[[[[1, 2, 3]]], [[[4, 5, 6]]], [[[7, 8, 9]]], [[[10, 11, 12]]]]
(3) For the following input of shape [1, 4, 4, 1]
, block_shape = [2, 2]
, and
paddings = [[0, 0], [0, 0]]
:
x = [[[[1], [2], [3], [4]],
[[5], [6], [7], [8]],
[[9], [10], [11], [12]],
[[13], [14], [15], [16]]]]
The output tensor has shape [4, 2, 2, 1]
and value:
x = [[[[1], [3]], [[9], [11]]],
[[[2], [4]], [[10], [12]]],
[[[5], [7]], [[13], [15]]],
[[[6], [8]], [[14], [16]]]]
(4) For the following input of shape [2, 2, 4, 1]
, block_shape = [2, 2]
, and
paddings = [[0, 0], [2, 0]]
:
x = [[[[1], [2], [3], [4]],
[[5], [6], [7], [8]]],
[[[9], [10], [11], [12]],
[[13], [14], [15], [16]]]]
The output tensor has shape [8, 1, 3, 1]
and value:
x = [[[[0], [1], [3]]], [[[0], [9], [11]]],
[[[0], [2], [4]]], [[[0], [10], [12]]],
[[[0], [5], [7]]], [[[0], [13], [15]]],
[[[0], [6], [8]]], [[[0], [14], [16]]]]
Among others, this operation is useful for reducing atrous convolution into regular convolution.
Returns | |
---|---|
A Tensor . Has the same type as input .
|