|
1
|
|
|
import torch |
|
2
|
|
|
import torchio as tio |
|
3
|
|
|
from ...utils import TorchioTestCase |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
class TestRandomAffine(TorchioTestCase): |
|
7
|
|
|
"""Tests for `RandomAffine`.""" |
|
8
|
|
|
def setUp(self): |
|
9
|
|
|
# Set image origin far from center |
|
10
|
|
|
super().setUp() |
|
11
|
|
|
affine = self.sample_subject.t1.affine |
|
12
|
|
|
affine[:3, 3] = 1e5 |
|
13
|
|
|
|
|
14
|
|
|
def test_rotation_image(self): |
|
15
|
|
|
# Rotation around image center |
|
16
|
|
|
transform = tio.RandomAffine( |
|
17
|
|
|
degrees=(90, 90), |
|
18
|
|
|
default_pad_value=0, |
|
19
|
|
|
center='image', |
|
20
|
|
|
) |
|
21
|
|
|
transformed = transform(self.sample_subject) |
|
22
|
|
|
total = transformed.t1.data.sum() |
|
23
|
|
|
self.assertNotEqual(total, 0) |
|
24
|
|
|
|
|
25
|
|
|
def test_rotation_origin(self): |
|
26
|
|
|
# Rotation around far away point, image should be empty |
|
27
|
|
|
transform = tio.RandomAffine( |
|
28
|
|
|
degrees=(90, 90), |
|
29
|
|
|
default_pad_value=0, |
|
30
|
|
|
center='origin', |
|
31
|
|
|
) |
|
32
|
|
|
transformed = transform(self.sample_subject) |
|
33
|
|
|
total = transformed.t1.data.sum() |
|
34
|
|
|
self.assertEqual(total, 0) |
|
35
|
|
|
|
|
36
|
|
|
def test_no_rotation(self): |
|
37
|
|
|
transform = tio.RandomAffine( |
|
38
|
|
|
scales=(1, 1), |
|
39
|
|
|
degrees=(0, 0), |
|
40
|
|
|
default_pad_value=0, |
|
41
|
|
|
center='image', |
|
42
|
|
|
) |
|
43
|
|
|
transformed = transform(self.sample_subject) |
|
44
|
|
|
self.assertTensorAlmostEqual( |
|
45
|
|
|
self.sample_subject.t1.data, |
|
46
|
|
|
transformed.t1.data, |
|
47
|
|
|
) |
|
48
|
|
|
|
|
49
|
|
|
transform = tio.RandomAffine( |
|
50
|
|
|
scales=(1, 1), |
|
51
|
|
|
degrees=(180, 180), |
|
52
|
|
|
default_pad_value=0, |
|
53
|
|
|
center='image', |
|
54
|
|
|
) |
|
55
|
|
|
transformed = transform(self.sample_subject) |
|
56
|
|
|
transformed = transform(transformed) |
|
57
|
|
|
self.assertTensorAlmostEqual( |
|
58
|
|
|
self.sample_subject.t1.data, |
|
59
|
|
|
transformed.t1.data, |
|
60
|
|
|
) |
|
61
|
|
|
|
|
62
|
|
|
def test_isotropic(self): |
|
63
|
|
|
tio.RandomAffine(isotropic=True)(self.sample_subject) |
|
64
|
|
|
|
|
65
|
|
|
def test_mean(self): |
|
66
|
|
|
tio.RandomAffine(default_pad_value='mean')(self.sample_subject) |
|
67
|
|
|
|
|
68
|
|
|
def test_otsu(self): |
|
69
|
|
|
tio.RandomAffine(default_pad_value='otsu')(self.sample_subject) |
|
70
|
|
|
|
|
71
|
|
|
def test_bad_center(self): |
|
72
|
|
|
with self.assertRaises(ValueError): |
|
73
|
|
|
tio.RandomAffine(center='bad') |
|
74
|
|
|
|
|
75
|
|
|
def test_negative_scales(self): |
|
76
|
|
|
with self.assertRaises(ValueError): |
|
77
|
|
|
tio.RandomAffine(scales=(-1, 1)) |
|
78
|
|
|
|
|
79
|
|
|
def test_scale_too_large(self): |
|
80
|
|
|
with self.assertRaises(ValueError): |
|
81
|
|
|
tio.RandomAffine(scales=1.5) |
|
82
|
|
|
|
|
83
|
|
|
def test_scales_range_with_negative_min(self): |
|
84
|
|
|
with self.assertRaises(ValueError): |
|
85
|
|
|
tio.RandomAffine(scales=(-1, 4)) |
|
86
|
|
|
|
|
87
|
|
|
def test_wrong_scales_type(self): |
|
88
|
|
|
with self.assertRaises(ValueError): |
|
89
|
|
|
tio.RandomAffine(scales='wrong') |
|
90
|
|
|
|
|
91
|
|
|
def test_wrong_degrees_type(self): |
|
92
|
|
|
with self.assertRaises(ValueError): |
|
93
|
|
|
tio.RandomAffine(degrees='wrong') |
|
94
|
|
|
|
|
95
|
|
|
def test_too_many_translation_values(self): |
|
96
|
|
|
with self.assertRaises(ValueError): |
|
97
|
|
|
tio.RandomAffine(translation=(-10, 4, 42)) |
|
98
|
|
|
|
|
99
|
|
|
def test_wrong_translation_type(self): |
|
100
|
|
|
with self.assertRaises(ValueError): |
|
101
|
|
|
tio.RandomAffine(translation='wrong') |
|
102
|
|
|
|
|
103
|
|
|
def test_wrong_center(self): |
|
104
|
|
|
with self.assertRaises(ValueError): |
|
105
|
|
|
tio.RandomAffine(center=0) |
|
106
|
|
|
|
|
107
|
|
|
def test_wrong_default_pad_value(self): |
|
108
|
|
|
with self.assertRaises(ValueError): |
|
109
|
|
|
tio.RandomAffine(default_pad_value='wrong') |
|
110
|
|
|
|
|
111
|
|
|
def test_wrong_image_interpolation_type(self): |
|
112
|
|
|
with self.assertRaises(TypeError): |
|
113
|
|
|
tio.RandomAffine(image_interpolation=0) |
|
114
|
|
|
|
|
115
|
|
|
def test_wrong_image_interpolation_value(self): |
|
116
|
|
|
with self.assertRaises(ValueError): |
|
117
|
|
|
tio.RandomAffine(image_interpolation='wrong') |
|
118
|
|
|
|
|
119
|
|
|
def test_incompatible_args_isotropic(self): |
|
120
|
|
|
with self.assertRaises(ValueError): |
|
121
|
|
|
tio.RandomAffine(scales=(0.8, 0.5, 0.1), isotropic=True) |
|
122
|
|
|
|
|
123
|
|
|
def test_parse_scales(self): |
|
124
|
|
|
def do_assert(transform): |
|
125
|
|
|
self.assertEqual(transform.scales, 3 * (0.9, 1.1)) |
|
126
|
|
|
do_assert(tio.RandomAffine(scales=0.1)) |
|
127
|
|
|
do_assert(tio.RandomAffine(scales=(0.9, 1.1))) |
|
128
|
|
|
do_assert(tio.RandomAffine(scales=3 * (0.1,))) |
|
129
|
|
|
do_assert(tio.RandomAffine(scales=3 * [0.9, 1.1])) |
|
130
|
|
|
|
|
131
|
|
|
def test_parse_degrees(self): |
|
132
|
|
|
def do_assert(transform): |
|
133
|
|
|
self.assertEqual(transform.degrees, 3 * (-10, 10)) |
|
134
|
|
|
do_assert(tio.RandomAffine(degrees=10)) |
|
135
|
|
|
do_assert(tio.RandomAffine(degrees=(-10, 10))) |
|
136
|
|
|
do_assert(tio.RandomAffine(degrees=3 * (10,))) |
|
137
|
|
|
do_assert(tio.RandomAffine(degrees=3 * [-10, 10])) |
|
138
|
|
|
|
|
139
|
|
|
def test_parse_translation(self): |
|
140
|
|
|
def do_assert(transform): |
|
141
|
|
|
self.assertEqual(transform.translation, 3 * (-10, 10)) |
|
142
|
|
|
do_assert(tio.RandomAffine(translation=10)) |
|
143
|
|
|
do_assert(tio.RandomAffine(translation=(-10, 10))) |
|
144
|
|
|
do_assert(tio.RandomAffine(translation=3 * (10,))) |
|
145
|
|
|
do_assert(tio.RandomAffine(translation=3 * [-10, 10])) |
|
146
|
|
|
|
|
147
|
|
|
def test_default_value_label_map(self): |
|
148
|
|
|
# From https://github.com/fepegar/torchio/issues/626 |
|
149
|
|
|
a = torch.tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]]).reshape(1, 3, 3, 1) |
|
150
|
|
|
image = tio.LabelMap(tensor=a) |
|
151
|
|
|
aff = tio.RandomAffine(translation=(0, 1, 1), default_pad_value='otsu') |
|
152
|
|
|
transformed = aff(image) |
|
153
|
|
|
assert all(n in (0, 1) for n in transformed.data.flatten()) |
|
154
|
|
|
|
|
155
|
|
|
def test_no_inverse(self): |
|
156
|
|
|
tensor = torch.zeros((1, 2, 2, 2)) |
|
157
|
|
|
tensor[0, 1, 1, 1] = 1 # most RAS voxel |
|
158
|
|
|
expected = torch.zeros((1, 2, 2, 2)) |
|
159
|
|
|
expected[0, 0, 1, 1] = 1 |
|
160
|
|
|
scales = 1, 1, 1 |
|
161
|
|
|
degrees = 0, 0, 90 # anterior should go left |
|
162
|
|
|
translation = 0, 0, 0 |
|
163
|
|
|
apply_affine = tio.Affine( |
|
164
|
|
|
scales, |
|
165
|
|
|
degrees, |
|
166
|
|
|
translation, |
|
167
|
|
|
) |
|
168
|
|
|
transformed = apply_affine(tensor) |
|
169
|
|
|
self.assertTensorAlmostEqual(transformed, expected) |
|
170
|
|
|
|
fepegar /
torchio
