File size: 1,893 Bytes
db6a3b7 a6bbecf db6a3b7 a6bbecf db6a3b7 a6bbecf db6a3b7 a6bbecf db6a3b7 a6bbecf db6a3b7 a6bbecf db6a3b7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 |
import torch
def pixel_shuffle_3d(x: torch.Tensor, scale_factor: int) -> torch.Tensor:
"""
3D pixel shuffle.
"""
B, C, H, W, D = x.shape
C_ = C // scale_factor**3
x = x.reshape(B, C_, scale_factor, scale_factor, scale_factor, H, W, D)
x = x.permute(0, 1, 5, 2, 6, 3, 7, 4)
x = x.reshape(B, C_, H * scale_factor, W * scale_factor, D * scale_factor)
return x
def patchify(x: torch.Tensor, patch_size: int):
"""
Patchify a tensor.
Args:
x (torch.Tensor): (N, C, *spatial) tensor
patch_size (int): Patch size
"""
DIM = x.dim() - 2
for d in range(2, DIM + 2):
assert (
x.shape[d] % patch_size == 0
), f"Dimension {d} of input tensor must be divisible by patch size, got {x.shape[d]} and {patch_size}"
x = x.reshape(
*x.shape[:2],
*sum([[x.shape[d] // patch_size, patch_size] for d in range(2, DIM + 2)], []),
)
x = x.permute(
0, 1, *([2 * i + 3 for i in range(DIM)] + [2 * i + 2 for i in range(DIM)])
)
x = x.reshape(x.shape[0], x.shape[1] * (patch_size**DIM), *(x.shape[-DIM:]))
return x
def unpatchify(x: torch.Tensor, patch_size: int):
"""
Unpatchify a tensor.
Args:
x (torch.Tensor): (N, C, *spatial) tensor
patch_size (int): Patch size
"""
DIM = x.dim() - 2
assert (
x.shape[1] % (patch_size**DIM) == 0
), f"Second dimension of input tensor must be divisible by patch size to unpatchify, got {x.shape[1]} and {patch_size ** DIM}"
x = x.reshape(
x.shape[0],
x.shape[1] // (patch_size**DIM),
*([patch_size] * DIM),
*(x.shape[-DIM:]),
)
x = x.permute(0, 1, *(sum([[2 + DIM + i, 2 + i] for i in range(DIM)], [])))
x = x.reshape(
x.shape[0], x.shape[1], *[x.shape[2 + 2 * i] * patch_size for i in range(DIM)]
)
return x
|