|
1
|
|
|
import warnings |
|
2
|
|
|
from collections import defaultdict |
|
3
|
|
|
from typing import Tuple, Optional, List |
|
4
|
|
|
|
|
5
|
|
|
import torch |
|
6
|
|
|
|
|
7
|
|
|
from ....utils import to_tuple |
|
8
|
|
|
from ....torchio import DATA, TypeRangeFloat |
|
9
|
|
|
from ....data.subject import Subject |
|
10
|
|
|
from ... import IntensityTransform |
|
11
|
|
|
from .. import RandomTransform |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class RandomGamma(RandomTransform, IntensityTransform): |
|
15
|
|
|
r"""Randomly change contrast of an image by raising its values to the power |
|
16
|
|
|
:math:`\gamma`. |
|
17
|
|
|
|
|
18
|
|
|
Args: |
|
19
|
|
|
log_gamma: Tuple :math:`(a, b)` to compute the exponent |
|
20
|
|
|
:math:`\gamma = e ^ \beta`, |
|
21
|
|
|
where :math:`\beta \sim \mathcal{U}(a, b)`. |
|
22
|
|
|
If a single value :math:`d` is provided, then |
|
23
|
|
|
:math:`\beta \sim \mathcal{U}(-d, d)`. |
|
24
|
|
|
Negative and positive values for this argument perform gamma |
|
25
|
|
|
compression and expansion, respectively. |
|
26
|
|
|
See the `Gamma correction`_ Wikipedia entry for more information. |
|
27
|
|
|
p: Probability that this transform will be applied. |
|
28
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
29
|
|
|
|
|
30
|
|
|
.. _Gamma correction: https://en.wikipedia.org/wiki/Gamma_correction |
|
31
|
|
|
|
|
32
|
|
|
.. warning:: Fractional exponentiation of negative values is generally not |
|
33
|
|
|
well-defined for non-complex numbers. |
|
34
|
|
|
If negative values are found in the input image :math:`I`, |
|
35
|
|
|
the applied transform is :math:`\text{sign}(I) |I|^\gamma`, |
|
36
|
|
|
instead of the usual :math:`I^\gamma`. The |
|
37
|
|
|
:py:class:`~torchio.transforms.preprocessing.intensity.rescale.RescaleIntensity` |
|
38
|
|
|
transform may be used to ensure that all values are positive. |
|
39
|
|
|
|
|
40
|
|
|
Example: |
|
41
|
|
|
>>> import torchio as tio |
|
42
|
|
|
>>> from torchio import RandomGamma |
|
43
|
|
|
>>> from tio.datasets import FPG |
|
44
|
|
|
>>> subject = FPG() |
|
45
|
|
|
>>> transform = RandomGamma(log_gamma=(-0.3, 0.3)) # gamma between 0.74 and 1.34 |
|
46
|
|
|
>>> transformed = transform(subject) |
|
47
|
|
|
""" |
|
48
|
|
|
def __init__( |
|
49
|
|
|
self, |
|
50
|
|
|
log_gamma: TypeRangeFloat = (-0.3, 0.3), |
|
51
|
|
|
p: float = 1, |
|
52
|
|
|
keys: Optional[List[str]] = None, |
|
53
|
|
|
): |
|
54
|
|
|
super().__init__(p=p, keys=keys) |
|
55
|
|
|
self.log_gamma_range = self.parse_range(log_gamma, 'log_gamma') |
|
56
|
|
|
|
|
57
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
58
|
|
|
arguments = defaultdict(dict) |
|
59
|
|
|
for name, image in self.get_images_dict(subject).items(): |
|
60
|
|
|
gammas = [self.get_params(self.log_gamma_range) for _ in image.data] |
|
61
|
|
|
arguments['gamma'][name] = gammas |
|
62
|
|
|
transform = Gamma(**arguments) |
|
63
|
|
|
transformed = transform(subject) |
|
64
|
|
|
return transformed |
|
65
|
|
|
|
|
66
|
|
|
@staticmethod |
|
67
|
|
|
def get_params(log_gamma_range: Tuple[float, float]) -> float: |
|
68
|
|
|
gamma = torch.FloatTensor(1).uniform_(*log_gamma_range).exp().item() |
|
69
|
|
|
return gamma |
|
70
|
|
|
|
|
71
|
|
|
|
|
72
|
|
|
class Gamma(IntensityTransform): |
|
73
|
|
|
r"""Change contrast of an image by raising its values to the power |
|
74
|
|
|
:math:`\gamma`. |
|
75
|
|
|
|
|
76
|
|
|
Args: |
|
77
|
|
|
gamma: Exponent to which values in the image will be raised. |
|
78
|
|
|
Negative and positive values for this argument perform gamma |
|
79
|
|
|
compression and expansion, respectively. |
|
80
|
|
|
See the `Gamma correction`_ Wikipedia entry for more information. |
|
81
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
82
|
|
|
|
|
83
|
|
|
.. _Gamma correction: https://en.wikipedia.org/wiki/Gamma_correction |
|
84
|
|
|
|
|
85
|
|
|
.. warning:: Fractional exponentiation of negative values is generally not |
|
86
|
|
|
well-defined for non-complex numbers. |
|
87
|
|
|
If negative values are found in the input image :math:`I`, |
|
88
|
|
|
the applied transform is :math:`\text{sign}(I) |I|^\gamma`, |
|
89
|
|
|
instead of the usual :math:`I^\gamma`. The |
|
90
|
|
|
:py:class:`~torchio.transforms.preprocessing.intensity.rescale.RescaleIntensity` |
|
91
|
|
|
transform may be used to ensure that all values are positive. |
|
92
|
|
|
|
|
93
|
|
|
Example: |
|
94
|
|
|
>>> import torchio as tio |
|
95
|
|
|
>>> from torchio import Gamma |
|
96
|
|
|
>>> from tio.datasets import FPG |
|
97
|
|
|
>>> subject = FPG() |
|
98
|
|
|
>>> transform = Gamma(0.8) |
|
99
|
|
|
>>> transformed = transform(subject) |
|
100
|
|
|
""" |
|
101
|
|
|
def __init__( |
|
102
|
|
|
self, |
|
103
|
|
|
gamma: float, |
|
104
|
|
|
keys: Optional[List[str]] = None, |
|
105
|
|
|
): |
|
106
|
|
|
super().__init__(keys=keys) |
|
107
|
|
|
self.gamma = gamma |
|
108
|
|
|
self.args_names = ('gamma',) |
|
109
|
|
|
self.invert_transform = False |
|
110
|
|
|
|
|
111
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
112
|
|
|
gamma = self.gamma |
|
113
|
|
|
for name, image in self.get_images_dict(subject).items(): |
|
114
|
|
|
if self.arguments_are_dict(): |
|
115
|
|
|
gamma = self.gamma[name] |
|
116
|
|
|
gammas = to_tuple(gamma, length=len(image.data)) |
|
117
|
|
|
transformed_tensors = [] |
|
118
|
|
|
for gamma, tensor in zip(gammas, image.data): |
|
119
|
|
|
if self.invert_transform: |
|
120
|
|
|
correction = power(tensor, 1 - gamma) |
|
121
|
|
|
transformed_tensor = tensor * correction |
|
122
|
|
|
else: |
|
123
|
|
|
transformed_tensor = power(tensor, gamma) |
|
124
|
|
|
transformed_tensors.append(transformed_tensor) |
|
125
|
|
|
image[DATA] = torch.stack(transformed_tensors) |
|
126
|
|
|
return subject |
|
127
|
|
|
|
|
128
|
|
|
|
|
129
|
|
|
def power(tensor, gamma): |
|
130
|
|
|
if tensor.min() < 0: |
|
131
|
|
|
message = ( |
|
132
|
|
|
'Negative values found in input tensor. See the documentation for' |
|
133
|
|
|
' more details on the implemented workaround:' |
|
134
|
|
|
' https://torchio.readthedocs.io/transforms/augmentation.html#randomgamma' |
|
135
|
|
|
) |
|
136
|
|
|
warnings.warn(message) |
|
137
|
|
|
output = tensor.sign() * tensor.abs() ** gamma |
|
138
|
|
|
else: |
|
139
|
|
|
output = tensor ** gamma |
|
140
|
|
|
return output |
|
141
|
|
|
|