Spaces:
Running
on
A10G
Running
on
A10G
File size: 11,036 Bytes
320e465 |
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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 |
from typing import List
import torch
import torch.nn.functional as F
from .kernels import normalize_kernel2d
def _compute_padding(kernel_size: List[int]) -> List[int]:
"""Compute padding tuple."""
# 4 or 6 ints: (padding_left, padding_right,padding_top,padding_bottom)
# https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad
if len(kernel_size) < 2:
raise AssertionError(kernel_size)
computed = [k - 1 for k in kernel_size]
# for even kernels we need to do asymmetric padding :(
out_padding = 2 * len(kernel_size) * [0]
for i in range(len(kernel_size)):
computed_tmp = computed[-(i + 1)]
pad_front = computed_tmp // 2
pad_rear = computed_tmp - pad_front
out_padding[2 * i + 0] = pad_front
out_padding[2 * i + 1] = pad_rear
return out_padding
def filter2d(
input: torch.Tensor,
kernel: torch.Tensor,
border_type: str = 'reflect',
normalized: bool = False,
padding: str = 'same',
) -> torch.Tensor:
r"""Convolve a tensor with a 2d kernel.
The function applies a given kernel to a tensor. The kernel is applied
independently at each depth channel of the tensor. Before applying the
kernel, the function applies padding according to the specified mode so
that the output remains in the same shape.
Args:
input: the input tensor with shape of
:math:`(B, C, H, W)`.
kernel: the kernel to be convolved with the input
tensor. The kernel shape must be :math:`(1, kH, kW)` or :math:`(B, kH, kW)`.
border_type: the padding mode to be applied before convolving.
The expected modes are: ``'constant'``, ``'reflect'``,
``'replicate'`` or ``'circular'``.
normalized: If True, kernel will be L1 normalized.
padding: This defines the type of padding.
2 modes available ``'same'`` or ``'valid'``.
Return:
torch.Tensor: the convolved tensor of same size and numbers of channels
as the input with shape :math:`(B, C, H, W)`.
Example:
>>> input = torch.tensor([[[
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 5., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],]]])
>>> kernel = torch.ones(1, 3, 3)
>>> filter2d(input, kernel, padding='same')
tensor([[[[0., 0., 0., 0., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 0., 0., 0., 0.]]]])
"""
if not isinstance(input, torch.Tensor):
raise TypeError(f"Input input is not torch.Tensor. Got {type(input)}")
if not isinstance(kernel, torch.Tensor):
raise TypeError(f"Input kernel is not torch.Tensor. Got {type(kernel)}")
if not isinstance(border_type, str):
raise TypeError(f"Input border_type is not string. Got {type(border_type)}")
if border_type not in ['constant', 'reflect', 'replicate', 'circular']:
raise ValueError(
f"Invalid border type, we expect 'constant', \
'reflect', 'replicate', 'circular'. Got:{border_type}"
)
if not isinstance(padding, str):
raise TypeError(f"Input padding is not string. Got {type(padding)}")
if padding not in ['valid', 'same']:
raise ValueError(f"Invalid padding mode, we expect 'valid' or 'same'. Got: {padding}")
if not len(input.shape) == 4:
raise ValueError(f"Invalid input shape, we expect BxCxHxW. Got: {input.shape}")
if (not len(kernel.shape) == 3) and not ((kernel.shape[0] == 0) or (kernel.shape[0] == input.shape[0])):
raise ValueError(f"Invalid kernel shape, we expect 1xHxW or BxHxW. Got: {kernel.shape}")
# prepare kernel
b, c, h, w = input.shape
tmp_kernel: torch.Tensor = kernel.unsqueeze(1).to(input)
if normalized:
tmp_kernel = normalize_kernel2d(tmp_kernel)
tmp_kernel = tmp_kernel.expand(-1, c, -1, -1)
height, width = tmp_kernel.shape[-2:]
# pad the input tensor
if padding == 'same':
padding_shape: List[int] = _compute_padding([height, width])
input = F.pad(input, padding_shape, mode=border_type)
# kernel and input tensor reshape to align element-wise or batch-wise params
tmp_kernel = tmp_kernel.reshape(-1, 1, height, width)
input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1))
# convolve the tensor with the kernel.
output = F.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1)
if padding == 'same':
out = output.view(b, c, h, w)
else:
out = output.view(b, c, h - height + 1, w - width + 1)
return out
def filter2d_separable(
input: torch.Tensor,
kernel_x: torch.Tensor,
kernel_y: torch.Tensor,
border_type: str = 'reflect',
normalized: bool = False,
padding: str = 'same',
) -> torch.Tensor:
r"""Convolve a tensor with two 1d kernels, in x and y directions.
The function applies a given kernel to a tensor. The kernel is applied
independently at each depth channel of the tensor. Before applying the
kernel, the function applies padding according to the specified mode so
that the output remains in the same shape.
Args:
input: the input tensor with shape of
:math:`(B, C, H, W)`.
kernel_x: the kernel to be convolved with the input
tensor. The kernel shape must be :math:`(1, kW)` or :math:`(B, kW)`.
kernel_y: the kernel to be convolved with the input
tensor. The kernel shape must be :math:`(1, kH)` or :math:`(B, kH)`.
border_type: the padding mode to be applied before convolving.
The expected modes are: ``'constant'``, ``'reflect'``,
``'replicate'`` or ``'circular'``.
normalized: If True, kernel will be L1 normalized.
padding: This defines the type of padding.
2 modes available ``'same'`` or ``'valid'``.
Return:
torch.Tensor: the convolved tensor of same size and numbers of channels
as the input with shape :math:`(B, C, H, W)`.
Example:
>>> input = torch.tensor([[[
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 5., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],]]])
>>> kernel = torch.ones(1, 3)
>>> filter2d_separable(input, kernel, kernel, padding='same')
tensor([[[[0., 0., 0., 0., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 0., 0., 0., 0.]]]])
"""
out_x = filter2d(input, kernel_x.unsqueeze(0), border_type, normalized, padding)
out = filter2d(out_x, kernel_y.unsqueeze(-1), border_type, normalized, padding)
return out
def filter3d(
input: torch.Tensor, kernel: torch.Tensor, border_type: str = 'replicate', normalized: bool = False
) -> torch.Tensor:
r"""Convolve a tensor with a 3d kernel.
The function applies a given kernel to a tensor. The kernel is applied
independently at each depth channel of the tensor. Before applying the
kernel, the function applies padding according to the specified mode so
that the output remains in the same shape.
Args:
input: the input tensor with shape of
:math:`(B, C, D, H, W)`.
kernel: the kernel to be convolved with the input
tensor. The kernel shape must be :math:`(1, kD, kH, kW)` or :math:`(B, kD, kH, kW)`.
border_type: the padding mode to be applied before convolving.
The expected modes are: ``'constant'``,
``'replicate'`` or ``'circular'``.
normalized: If True, kernel will be L1 normalized.
Return:
the convolved tensor of same size and numbers of channels
as the input with shape :math:`(B, C, D, H, W)`.
Example:
>>> input = torch.tensor([[[
... [[0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.]],
... [[0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 5., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.]],
... [[0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.],
... [0., 0., 0., 0., 0.]]
... ]]])
>>> kernel = torch.ones(1, 3, 3, 3)
>>> filter3d(input, kernel)
tensor([[[[[0., 0., 0., 0., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 0., 0., 0., 0.]],
<BLANKLINE>
[[0., 0., 0., 0., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 0., 0., 0., 0.]],
<BLANKLINE>
[[0., 0., 0., 0., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 5., 5., 5., 0.],
[0., 0., 0., 0., 0.]]]]])
"""
if not isinstance(input, torch.Tensor):
raise TypeError(f"Input border_type is not torch.Tensor. Got {type(input)}")
if not isinstance(kernel, torch.Tensor):
raise TypeError(f"Input border_type is not torch.Tensor. Got {type(kernel)}")
if not isinstance(border_type, str):
raise TypeError(f"Input border_type is not string. Got {type(kernel)}")
if not len(input.shape) == 5:
raise ValueError(f"Invalid input shape, we expect BxCxDxHxW. Got: {input.shape}")
if not len(kernel.shape) == 4 and kernel.shape[0] != 1:
raise ValueError(f"Invalid kernel shape, we expect 1xDxHxW. Got: {kernel.shape}")
# prepare kernel
b, c, d, h, w = input.shape
tmp_kernel: torch.Tensor = kernel.unsqueeze(1).to(input)
if normalized:
bk, dk, hk, wk = kernel.shape
tmp_kernel = normalize_kernel2d(tmp_kernel.view(bk, dk, hk * wk)).view_as(tmp_kernel)
tmp_kernel = tmp_kernel.expand(-1, c, -1, -1, -1)
# pad the input tensor
depth, height, width = tmp_kernel.shape[-3:]
padding_shape: List[int] = _compute_padding([depth, height, width])
input_pad: torch.Tensor = F.pad(input, padding_shape, mode=border_type)
# kernel and input tensor reshape to align element-wise or batch-wise params
tmp_kernel = tmp_kernel.reshape(-1, 1, depth, height, width)
input_pad = input_pad.view(-1, tmp_kernel.size(0), input_pad.size(-3), input_pad.size(-2), input_pad.size(-1))
# convolve the tensor with the kernel.
output = F.conv3d(input_pad, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1)
return output.view(b, c, d, h, w) |