|
1
|
|
|
from typing import Union, Sequence |
|
2
|
|
|
|
|
3
|
|
|
import torch |
|
4
|
|
|
import numpy as np |
|
5
|
|
|
from torchvision.transforms import Compose as PyTorchCompose |
|
6
|
|
|
|
|
7
|
|
|
from ...data.subject import Subject |
|
8
|
|
|
from .. import Transform |
|
9
|
|
|
from . import RandomTransform |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class Compose(Transform): |
|
13
|
|
|
"""Compose several transforms together. |
|
14
|
|
|
|
|
15
|
|
|
Args: |
|
16
|
|
|
transforms: Sequence of instances of |
|
17
|
|
|
:py:class:`~torchio.transforms.transform.Transform`. |
|
18
|
|
|
p: Probability that this transform will be applied. |
|
19
|
|
|
|
|
20
|
|
|
.. note:: |
|
21
|
|
|
This is a thin wrapper of :py:class:`torchvision.transforms.Compose`. |
|
22
|
|
|
""" |
|
23
|
|
|
def __init__(self, transforms: Sequence[Transform], p: float = 1): |
|
24
|
|
|
super().__init__(p=p) |
|
25
|
|
|
self.transform = PyTorchCompose(transforms) |
|
26
|
|
|
self.transforms = self.transform.transforms |
|
27
|
|
|
|
|
28
|
|
|
def __len__(self): |
|
29
|
|
|
return len(self.transforms) |
|
30
|
|
|
|
|
31
|
|
|
def __getitem__(self, index): |
|
32
|
|
|
return self.transforms[index] |
|
33
|
|
|
|
|
34
|
|
|
def __repr__(self): |
|
35
|
|
|
return self.transform.__repr__() |
|
36
|
|
|
|
|
37
|
|
|
def apply_transform(self, subject: Subject): |
|
38
|
|
|
return self.transform(subject) |
|
39
|
|
|
|
|
40
|
|
|
def is_invertible(self): |
|
41
|
|
|
return all(t.is_invertible() for t in self.transforms) |
|
42
|
|
|
|
|
43
|
|
|
def inverse(self): |
|
44
|
|
|
transforms = [ |
|
45
|
|
|
transform.inverse() |
|
46
|
|
|
for transform in self.transforms |
|
47
|
|
|
if transform.is_invertible() |
|
48
|
|
|
] |
|
49
|
|
|
return Compose(reversed(transforms)) |
|
50
|
|
|
|
|
51
|
|
|
|
|
52
|
|
|
class OneOf(RandomTransform): |
|
53
|
|
|
"""Apply only one of the given transforms. |
|
54
|
|
|
|
|
55
|
|
|
Args: |
|
56
|
|
|
transforms: Dictionary with instances of |
|
57
|
|
|
:py:class:`~torchio.transforms.transform.Transform` as keys and |
|
58
|
|
|
probabilities as values. Probabilities are normalized so they sum |
|
59
|
|
|
to one. If a sequence is given, the same probability will be |
|
60
|
|
|
assigned to each transform. |
|
61
|
|
|
p: Probability that this transform will be applied. |
|
62
|
|
|
|
|
63
|
|
|
Example: |
|
64
|
|
|
>>> import torchio as tio |
|
65
|
|
|
>>> colin = tio.datasets.Colin27() |
|
66
|
|
|
>>> transforms_dict = { |
|
67
|
|
|
... tio.RandomAffine(): 0.75, |
|
68
|
|
|
... tio.RandomElasticDeformation(): 0.25, |
|
69
|
|
|
... } # Using 3 and 1 as probabilities would have the same effect |
|
70
|
|
|
>>> transform = torchio.transforms.OneOf(transforms_dict) |
|
71
|
|
|
>>> transformed = transform(colin) |
|
72
|
|
|
|
|
73
|
|
|
""" |
|
74
|
|
|
def __init__( |
|
75
|
|
|
self, |
|
76
|
|
|
transforms: Union[dict, Sequence[Transform]], |
|
77
|
|
|
p: float = 1, |
|
78
|
|
|
): |
|
79
|
|
|
super().__init__(p=p) |
|
80
|
|
|
self.transforms_dict = self._get_transforms_dict(transforms) |
|
81
|
|
|
|
|
82
|
|
|
def apply_transform(self, subject: Subject): |
|
83
|
|
|
weights = torch.Tensor(list(self.transforms_dict.values())) |
|
84
|
|
|
index = torch.multinomial(weights, 1) |
|
85
|
|
|
transforms = list(self.transforms_dict.keys()) |
|
86
|
|
|
transform = transforms[index] |
|
87
|
|
|
transformed = transform(subject) |
|
88
|
|
|
return transformed |
|
89
|
|
|
|
|
90
|
|
|
def _get_transforms_dict(self, transforms: Union[dict, Sequence]): |
|
91
|
|
|
if isinstance(transforms, dict): |
|
92
|
|
|
transforms_dict = dict(transforms) |
|
93
|
|
|
self._normalize_probabilities(transforms_dict) |
|
94
|
|
|
else: |
|
95
|
|
|
try: |
|
96
|
|
|
p = 1 / len(transforms) |
|
97
|
|
|
except TypeError as e: |
|
98
|
|
|
message = ( |
|
99
|
|
|
'Transforms argument must be a dictionary or a sequence,' |
|
100
|
|
|
f' not {type(transforms)}' |
|
101
|
|
|
) |
|
102
|
|
|
raise ValueError(message) from e |
|
103
|
|
|
transforms_dict = {transform: p for transform in transforms} |
|
104
|
|
|
for transform in transforms_dict: |
|
105
|
|
|
if not isinstance(transform, Transform): |
|
106
|
|
|
message = ( |
|
107
|
|
|
'All keys in transform_dict must be instances of' |
|
108
|
|
|
f'torchio.Transform, not "{type(transform)}"' |
|
109
|
|
|
) |
|
110
|
|
|
raise ValueError(message) |
|
111
|
|
|
return transforms_dict |
|
112
|
|
|
|
|
113
|
|
|
@staticmethod |
|
114
|
|
|
def _normalize_probabilities(transforms_dict: dict): |
|
115
|
|
|
probabilities = np.array(list(transforms_dict.values()), dtype=float) |
|
116
|
|
|
if np.any(probabilities < 0): |
|
117
|
|
|
message = ( |
|
118
|
|
|
'Probabilities must be greater or equal to zero,' |
|
119
|
|
|
f' not "{probabilities}"' |
|
120
|
|
|
) |
|
121
|
|
|
raise ValueError(message) |
|
122
|
|
|
if np.all(probabilities == 0): |
|
123
|
|
|
message = ( |
|
124
|
|
|
'At least one probability must be greater than zero,' |
|
125
|
|
|
f' but they are "{probabilities}"' |
|
126
|
|
|
) |
|
127
|
|
|
raise ValueError(message) |
|
128
|
|
|
for transform, probability in transforms_dict.items(): |
|
129
|
|
|
transforms_dict[transform] = probability / probabilities.sum() |
|
130
|
|
|
|
fepegar /
torchio
