pixel_shuffle¶

Rearranges elements in a tensor of shape \((*, C \cdot r^2, H, W)\) into shape \((*, C, H \cdot r, W \cdot r)\), where \(r\) is the upscale factor. The input’s channel dimension must be divisible by \(r^2\).

ATen source: aten.pixel_shuffle (preserved as composite by get_decomp_table())

Inputs¶

Name	Description
`input`	Tensor with at least 3 dims; size of dim `-3` must be divisible by `upscale_factor ** 2`

Attributes¶

Name	Type	Description
`upscale_factor`	`int`	Factor by which to increase spatial resolution
`version`	`int`	Composite op version

Output¶

Name	Description
`output`	Same leading dims; channel divided by \(r^2\), spatial dims multiplied by \(r\)

PyTorch example¶

import torch

N, C, H, W = 8, 1000, 32, 32
input = torch.randn(N, C, H, W)
# Output shape: (8, 10, 320, 320)
output = torch.nn.functional.pixel_shuffle(input, upscale_factor=10)

Reference¶

torch.nn.PixelShuffle