|
1
|
|
|
import copy |
|
2
|
|
|
from collections import defaultdict |
|
3
|
|
|
from typing import Tuple, Optional, Union, List, Dict, Sequence |
|
4
|
|
|
|
|
5
|
|
|
import torch |
|
6
|
|
|
from ....torchio import DATA |
|
7
|
|
|
from ....utils import get_random_seed |
|
8
|
|
|
from ....data.subject import Subject |
|
9
|
|
|
from ... import IntensityTransform |
|
10
|
|
|
from .. import RandomTransform |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class RandomNoise(RandomTransform, IntensityTransform): |
|
14
|
|
|
r"""Add Gaussian noise with random parameters. |
|
15
|
|
|
|
|
16
|
|
|
Add noise sampled from a normal distribution with random parameters. |
|
17
|
|
|
|
|
18
|
|
|
Args: |
|
19
|
|
|
mean: Mean :math:`\mu` of the Gaussian distribution |
|
20
|
|
|
from which the noise is sampled. |
|
21
|
|
|
If two values :math:`(a, b)` are provided, |
|
22
|
|
|
then :math:`\mu \sim \mathcal{U}(a, b)`. |
|
23
|
|
|
If only one value :math:`d` is provided, |
|
24
|
|
|
:math:`\mu \sim \mathcal{U}(-d, d)`. |
|
25
|
|
|
std: Standard deviation :math:`\sigma` of the Gaussian distribution |
|
26
|
|
|
from which the noise is sampled. |
|
27
|
|
|
If two values :math:`(a, b)` are provided, |
|
28
|
|
|
then :math:`\sigma \sim \mathcal{U}(a, b)`. |
|
29
|
|
|
If only one value :math:`d` is provided, |
|
30
|
|
|
:math:`\sigma \sim \mathcal{U}(0, d)`. |
|
31
|
|
|
p: Probability that this transform will be applied. |
|
32
|
|
|
seed: See :py:class:`~torchio.transforms.augmentation.RandomTransform`. |
|
33
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
34
|
|
|
""" |
|
35
|
|
|
def __init__( |
|
36
|
|
|
self, |
|
37
|
|
|
mean: Union[float, Tuple[float, float]] = 0, |
|
38
|
|
|
std: Union[float, Tuple[float, float]] = (0, 0.25), |
|
39
|
|
|
p: float = 1, |
|
40
|
|
|
keys: Optional[Sequence[str]] = None, |
|
41
|
|
|
): |
|
42
|
|
|
super().__init__(p=p, keys=keys) |
|
43
|
|
|
self.mean_range = self.parse_range(mean, 'mean') |
|
44
|
|
|
self.std_range = self.parse_range(std, 'std', min_constraint=0) |
|
45
|
|
|
|
|
46
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
47
|
|
|
arguments = defaultdict(dict) |
|
48
|
|
|
for image_name in self.get_images_dict(subject): |
|
49
|
|
|
mean, std, seed = self.get_params(self.mean_range, self.std_range) |
|
50
|
|
|
arguments['mean'][image_name] = mean |
|
51
|
|
|
arguments['std'][image_name] = std |
|
52
|
|
|
arguments['seed'][image_name] = seed |
|
53
|
|
|
transform = Noise(**arguments) |
|
54
|
|
|
transformed = transform(subject) |
|
55
|
|
|
return transformed |
|
56
|
|
|
|
|
57
|
|
|
@staticmethod |
|
58
|
|
|
def get_params( |
|
59
|
|
|
mean_range: Tuple[float, float], |
|
60
|
|
|
std_range: Tuple[float, float], |
|
61
|
|
|
) -> Tuple[float, float]: |
|
62
|
|
|
mean = torch.FloatTensor(1).uniform_(*mean_range).item() |
|
63
|
|
|
std = torch.FloatTensor(1).uniform_(*std_range).item() |
|
64
|
|
|
seed = get_random_seed() |
|
65
|
|
|
return mean, std, seed |
|
66
|
|
|
|
|
67
|
|
|
|
|
68
|
|
|
class Noise(IntensityTransform): |
|
69
|
|
|
r"""Add Gaussian noise. |
|
70
|
|
|
|
|
71
|
|
|
Add noise sampled from a normal distribution. |
|
72
|
|
|
|
|
73
|
|
|
Args: |
|
74
|
|
|
mean: Mean :math:`\mu` of the Gaussian distribution |
|
75
|
|
|
from which the noise is sampled. |
|
76
|
|
|
std: Standard deviation :math:`\sigma` of the Gaussian distribution |
|
77
|
|
|
from which the noise is sampled. |
|
78
|
|
|
seed: |
|
79
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
80
|
|
|
""" |
|
81
|
|
|
def __init__( |
|
82
|
|
|
self, |
|
83
|
|
|
mean: Union[float, Dict[str, float]], |
|
84
|
|
|
std: Union[float, Dict[str, float]], |
|
85
|
|
|
seed: Union[int, Sequence[int]], |
|
86
|
|
|
keys: Optional[List[str]] = None, |
|
87
|
|
|
): |
|
88
|
|
|
super().__init__(keys=keys) |
|
89
|
|
|
self.mean = mean |
|
90
|
|
|
self.std = std |
|
91
|
|
|
self.seed = seed |
|
92
|
|
|
self._has_dicts = self.arguments_are_dict() |
|
93
|
|
|
self.invert_transform = False |
|
94
|
|
|
self.args_names = 'mean', 'std', 'seed' |
|
95
|
|
|
|
|
96
|
|
|
def arguments_are_dict(self): |
|
97
|
|
|
mean_dict = isinstance(self.mean, dict) |
|
98
|
|
|
std_dict = isinstance(self.std, dict) |
|
99
|
|
|
seed_dict = isinstance(self.seed, dict) |
|
100
|
|
|
three_bools = mean_dict, std_dict, seed_dict |
|
101
|
|
|
if all(three_bools): |
|
102
|
|
|
return True |
|
103
|
|
|
elif not any(three_bools): |
|
104
|
|
|
return False |
|
105
|
|
|
else: |
|
106
|
|
|
message = 'All arguments must have the same type: float or dict' |
|
107
|
|
|
raise ValueError(message) |
|
108
|
|
|
|
|
109
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
110
|
|
|
args = self.mean, self.std, self.seed |
|
111
|
|
|
for name, image in self.get_images_dict(subject).items(): |
|
112
|
|
|
if self._has_dicts: |
|
113
|
|
|
args = self.mean[name], self.std[name], self.seed[name] |
|
114
|
|
|
noise = get_noise(image[DATA], *args) |
|
|
|
|
|
|
115
|
|
|
if self.invert_transform: |
|
116
|
|
|
noise *= -1 |
|
117
|
|
|
image[DATA] = image[DATA] + noise |
|
118
|
|
|
return subject |
|
119
|
|
|
|
|
120
|
|
|
def inverse(self): |
|
121
|
|
|
new = copy.deepcopy(self) |
|
122
|
|
|
new.invert_transform = not self.invert_transform |
|
123
|
|
|
return new |
|
124
|
|
|
|
|
125
|
|
|
|
|
126
|
|
|
def get_randn(shape: Sequence[int], seed: int) -> torch.Tensor: |
|
127
|
|
|
torch_rng_state = torch.random.get_rng_state() |
|
128
|
|
|
torch.manual_seed(seed) |
|
129
|
|
|
noise = torch.randn(*shape) |
|
130
|
|
|
torch.random.set_rng_state(torch_rng_state) |
|
131
|
|
|
return noise |
|
132
|
|
|
|
|
133
|
|
|
|
|
134
|
|
|
def get_noise( |
|
135
|
|
|
tensor: torch.Tensor, |
|
136
|
|
|
mean: float, |
|
137
|
|
|
std: float, |
|
138
|
|
|
seed: Optional[int] = None) -> torch.Tensor: |
|
139
|
|
|
return get_randn(tensor.shape, seed=seed) * std + mean |
|
140
|
|
|
|
fepegar /
torchio
Loading history...