File size: 8,640 Bytes
9c6594c |
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 |
# Copyright The Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import Optional, Union
import torch
from torch import Tensor, tensor
from torch.nn.functional import conv2d, pad
from typing_extensions import Literal
from torchmetrics.utilities.checks import _check_same_shape
from torchmetrics.utilities.distributed import reduce
def _scc_update(preds: Tensor, target: Tensor, hp_filter: Tensor, window_size: int) -> tuple[Tensor, Tensor, Tensor]:
"""Update and returns variables required to compute Spatial Correlation Coefficient.
Args:
preds: Predicted tensor
target: Ground truth tensor
hp_filter: High-pass filter tensor
window_size: Local window size integer
Return:
Tuple of (preds, target, hp_filter) tensors
Raises:
ValueError:
If ``preds`` and ``target`` have different number of channels
If ``preds`` and ``target`` have different shapes
If ``preds`` and ``target`` have invalid shapes
If ``window_size`` is not a positive integer
If ``window_size`` is greater than the size of the image
"""
if preds.dtype != target.dtype:
target = target.to(preds.dtype)
_check_same_shape(preds, target)
if preds.ndim not in (3, 4):
raise ValueError(
"Expected `preds` and `target` to have batch of colored images with BxCxHxW shape"
" or batch of grayscale images of BxHxW shape."
f" Got preds: {preds.shape} and target: {target.shape}."
)
if len(preds.shape) == 3:
preds = preds.unsqueeze(1)
target = target.unsqueeze(1)
if not window_size > 0:
raise ValueError(f"Expected `window_size` to be a positive integer. Got {window_size}.")
if window_size > preds.size(2) or window_size > preds.size(3):
raise ValueError(
f"Expected `window_size` to be less than or equal to the size of the image."
f" Got window_size: {window_size} and image size: {preds.size(2)}x{preds.size(3)}."
)
preds = preds.to(torch.float32)
target = target.to(torch.float32)
hp_filter = hp_filter[None, None, :].to(dtype=preds.dtype, device=preds.device)
return preds, target, hp_filter
def _symmetric_reflect_pad_2d(input_img: Tensor, pad: Union[int, tuple[int, ...]]) -> Tensor:
"""Applies symmetric padding to the 2D image tensor input using ``reflect`` mode (d c b a | a b c d | d c b a)."""
if isinstance(pad, int):
pad = (pad, pad, pad, pad)
if len(pad) != 4:
raise ValueError(f"Expected padding to have length 4, but got {len(pad)}")
left_pad = input_img[:, :, :, 0 : pad[0]].flip(dims=[3])
right_pad = input_img[:, :, :, -pad[1] :].flip(dims=[3])
padded = torch.cat([left_pad, input_img, right_pad], dim=3)
top_pad = padded[:, :, 0 : pad[2], :].flip(dims=[2])
bottom_pad = padded[:, :, -pad[3] :, :].flip(dims=[2])
return torch.cat([top_pad, padded, bottom_pad], dim=2)
def _signal_convolve_2d(input_img: Tensor, kernel: Tensor) -> Tensor:
"""Applies 2D signal convolution to the input tensor with the given kernel."""
left_padding = int(math.floor((kernel.size(3) - 1) / 2))
right_padding = int(math.ceil((kernel.size(3) - 1) / 2))
top_padding = int(math.floor((kernel.size(2) - 1) / 2))
bottom_padding = int(math.ceil((kernel.size(2) - 1) / 2))
padded = _symmetric_reflect_pad_2d(input_img, pad=(left_padding, right_padding, top_padding, bottom_padding))
kernel = kernel.flip([2, 3])
return conv2d(padded, kernel, stride=1, padding=0)
def _hp_2d_laplacian(input_img: Tensor, kernel: Tensor) -> Tensor:
"""Applies 2-D Laplace filter to the input tensor with the given high pass filter."""
return _signal_convolve_2d(input_img, kernel) * 2.0
def _local_variance_covariance(preds: Tensor, target: Tensor, window: Tensor) -> tuple[Tensor, Tensor, Tensor]:
"""Computes local variance and covariance of the input tensors."""
# This code is inspired by
# https://github.com/andrewekhalel/sewar/blob/master/sewar/full_ref.py#L187.
left_padding = int(math.ceil((window.size(3) - 1) / 2))
right_padding = int(math.floor((window.size(3) - 1) / 2))
preds = pad(preds, (left_padding, right_padding, left_padding, right_padding))
target = pad(target, (left_padding, right_padding, left_padding, right_padding))
preds_mean = conv2d(preds, window, stride=1, padding=0)
target_mean = conv2d(target, window, stride=1, padding=0)
preds_var = conv2d(preds**2, window, stride=1, padding=0) - preds_mean**2
target_var = conv2d(target**2, window, stride=1, padding=0) - target_mean**2
target_preds_cov = conv2d(target * preds, window, stride=1, padding=0) - target_mean * preds_mean
return preds_var, target_var, target_preds_cov
def _scc_per_channel_compute(preds: Tensor, target: Tensor, hp_filter: Tensor, window_size: int) -> Tensor:
"""Computes per channel Spatial Correlation Coefficient.
Args:
preds: estimated image of Bx1xHxW shape.
target: ground truth image of Bx1xHxW shape.
hp_filter: 2D high-pass filter.
window_size: size of window for local mean calculation.
Return:
Tensor with Spatial Correlation Coefficient score
"""
dtype = preds.dtype
device = preds.device
# This code is inspired by
# https://github.com/andrewekhalel/sewar/blob/master/sewar/full_ref.py#L187.
window = torch.ones(size=(1, 1, window_size, window_size), dtype=dtype, device=device) / (window_size**2)
preds_hp = _hp_2d_laplacian(preds, hp_filter)
target_hp = _hp_2d_laplacian(target, hp_filter)
preds_var, target_var, target_preds_cov = _local_variance_covariance(preds_hp, target_hp, window)
preds_var[preds_var < 0] = 0
target_var[target_var < 0] = 0
den = torch.sqrt(target_var) * torch.sqrt(preds_var)
idx = den == 0
den[den == 0] = 1
scc = target_preds_cov / den
scc[idx] = 0
return scc
def spatial_correlation_coefficient(
preds: Tensor,
target: Tensor,
hp_filter: Optional[Tensor] = None,
window_size: int = 8,
reduction: Optional[Literal["mean", "none", None]] = "mean",
) -> Tensor:
"""Compute Spatial Correlation Coefficient (SCC_).
Args:
preds: predicted images of shape ``(N,C,H,W)`` or ``(N,H,W)``.
target: ground truth images of shape ``(N,C,H,W)`` or ``(N,H,W)``.
hp_filter: High-pass filter tensor. default: tensor([[-1,-1,-1],[-1,8,-1],[-1,-1,-1]])
window_size: Local window size integer. default: 8,
reduction: Reduction method for output tensor. If ``None`` or ``"none"``,
returns a tensor with the per sample results. default: ``"mean"``.
Return:
Tensor with scc score
Example:
>>> from torch import randn
>>> from torchmetrics.functional.image import spatial_correlation_coefficient as scc
>>> x = randn(5, 3, 16, 16)
>>> scc(x, x)
tensor(1.)
>>> x = randn(5, 16, 16)
>>> scc(x, x)
tensor(1.)
>>> x = randn(5, 3, 16, 16)
>>> y = randn(5, 3, 16, 16)
>>> scc(x, y, reduction="none")
tensor([0.0223, 0.0256, 0.0616, 0.0159, 0.0170])
"""
if hp_filter is None:
hp_filter = tensor([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]])
if reduction is None:
reduction = "none"
if reduction not in ("mean", "none"):
raise ValueError(f"Expected reduction to be 'mean' or 'none', but got {reduction}")
preds, target, hp_filter = _scc_update(preds, target, hp_filter, window_size)
per_channel = [
_scc_per_channel_compute(
preds[:, i, :, :].unsqueeze(1), target[:, i, :, :].unsqueeze(1), hp_filter, window_size
)
for i in range(preds.size(1))
]
if reduction == "none":
return torch.mean(torch.cat(per_channel, dim=1), dim=[1, 2, 3])
if reduction == "mean":
return reduce(torch.cat(per_channel, dim=1), reduction="elementwise_mean")
return None
|