paddings = [[pad_top, pad_bottom], [pad_left, pad_right]]

The effective spatial dimensions of the zero-padded input tensor will be:

height_pad = pad_top + height + pad_bottom width_pad = pad_left + width + pad_right

The attr block_size must be greater than one. It indicates the block size.

• Non-overlapping blocks of size block_size x block size in the height and width dimensions are rearranged into the batch dimension at each location.
• The batch of the output tensor is batch * block_size * block_size.
• Both height_pad and width_pad must be divisible by block_size.

The shape of the output will be:

[batchblock_sizeblock_size, height_pad/block_size, width_pad/block_size, depth]

Some examples:

(1) For the following input of shape [1, 2, 2, 1] and block_size of 2:

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] and block_size of 2:

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] and block_size of 2:

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] and block_size of 2:

x = [[[[1],   [2],  [3],  [4]],
[[5],   [6],  [7],  [8]]],
[[[9],  [10], [11],  [12]],
[[13], [14], [15],  [16]]]]

The output tensor has shape [8, 1, 2, 1] and value:

x = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]],
[[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]]

Among others, this operation is useful for reducing atrous convolution into regular convolution.