|
1
|
|
|
import warnings |
|
2
|
|
|
import torch |
|
3
|
|
|
import torchio |
|
4
|
|
|
import numpy as np |
|
5
|
|
|
from torchio import Subject, ScalarImage, LabelMap |
|
6
|
|
|
from torchio.transforms import RandomNoise, compose_from_history, Compose, RandomSpike, OneOf |
|
7
|
|
|
from ..utils import TorchioTestCase |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class TestReproducibility(TorchioTestCase): |
|
11
|
|
|
|
|
12
|
|
|
def get_subjects(self): |
|
13
|
|
|
subject1 = Subject(img=ScalarImage(tensor=torch.ones(1, 4, 4, 4))) |
|
14
|
|
|
subject2 = Subject(img=ScalarImage(tensor=torch.ones(1, 4, 4, 4))) |
|
15
|
|
|
return subject1, subject2 |
|
16
|
|
|
|
|
17
|
|
|
def apply_transforms(self, subject, trsfm_list, seeds_list): |
|
18
|
|
|
s = subject |
|
19
|
|
|
for trsfm, seed in zip(trsfm_list, seeds_list): |
|
20
|
|
|
if seed: |
|
21
|
|
|
s = trsfm(s, seed=seed) |
|
22
|
|
|
else: |
|
23
|
|
|
s = trsfm(s) |
|
24
|
|
|
return s |
|
25
|
|
|
|
|
26
|
|
|
def random_stuff(self, seed=42): |
|
27
|
|
|
transform = RandomNoise(std=(100, 100)) |
|
28
|
|
|
transformed = transform(self.sample_subject, seed=seed) |
|
29
|
|
|
value = transformed.img.data.sum().item() |
|
30
|
|
|
seed = transformed.history[0][1]['seed'] |
|
31
|
|
|
return value, seed |
|
32
|
|
|
|
|
33
|
|
|
def test_rng_state(self): |
|
34
|
|
|
trsfm = RandomNoise() |
|
35
|
|
|
subject1, subject2 = self.get_subjects() |
|
36
|
|
|
transformed1 = trsfm(subject1) |
|
37
|
|
|
seed1 = transformed1.history[0][1]['seed'] |
|
38
|
|
|
value1_torch, value1_np = torch.rand(1).item(), np.random.rand() |
|
39
|
|
|
transformed2 = trsfm(subject2, seed=seed1) |
|
40
|
|
|
value2_torch, value2_np = torch.rand(1).item(), np.random.rand() |
|
41
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
42
|
|
|
self.assertNotEqual(value1_torch, value2_torch) |
|
43
|
|
|
self.assertNotEqual(value1_np, value2_np) |
|
44
|
|
|
self.assertTensorEqual(data1, data2) |
|
45
|
|
|
|
|
46
|
|
|
def test_reproducibility_seed(self): |
|
47
|
|
|
trsfm = RandomNoise() |
|
48
|
|
|
subject1, subject2 = self.get_subjects() |
|
49
|
|
|
transformed1 = trsfm(subject1) |
|
50
|
|
|
seed1 = transformed1.history[0][1]['seed'] |
|
51
|
|
|
transformed2 = trsfm(subject2, seed=seed1) |
|
52
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
53
|
|
|
seed2 = transformed2.history[0][1]['seed'] |
|
54
|
|
|
self.assertTensorEqual(data1, data2) |
|
55
|
|
|
self.assertEqual(seed1, seed2) |
|
56
|
|
|
|
|
57
|
|
|
def test_reproducibility_no_seed(self): |
|
58
|
|
|
trsfm = RandomNoise() |
|
59
|
|
|
subject1, subject2 = self.get_subjects() |
|
60
|
|
|
transformed1 = trsfm(subject1) |
|
61
|
|
|
transformed2 = trsfm(subject2) |
|
62
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
63
|
|
|
seed1, seed2 = transformed1.history[0][1]['seed'], transformed2.history[0][1]['seed'] |
|
64
|
|
|
self.assertNotEqual(seed1, seed2) |
|
65
|
|
|
self.assertTensorNotEqual(data1, data2) |
|
66
|
|
|
|
|
67
|
|
|
def test_reproducibility_from_history(self): |
|
68
|
|
|
trsfm = RandomNoise() |
|
69
|
|
|
subject1, subject2 = self.get_subjects() |
|
70
|
|
|
transformed1 = trsfm(subject1) |
|
71
|
|
|
history1 = transformed1.history |
|
72
|
|
|
trsfm_hist, seeds_hist = compose_from_history(history=history1) |
|
73
|
|
|
transformed2 = self.apply_transforms(subject2, trsfm_list=trsfm_hist, seeds_list=seeds_hist) |
|
74
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
75
|
|
|
self.assertTensorEqual(data1, data2) |
|
76
|
|
|
|
|
77
|
|
|
def test_reproducibility_compose(self): |
|
78
|
|
|
trsfm = Compose([RandomNoise(p=0.0), RandomSpike(num_spikes=3, p=1.0)]) |
|
79
|
|
|
subject1, subject2 = self.get_subjects() |
|
80
|
|
|
transformed1 = trsfm(subject1) |
|
81
|
|
|
history1 = transformed1.history |
|
82
|
|
|
trsfm_hist, seeds_hist = compose_from_history(history=history1) |
|
83
|
|
|
transformed2 = self.apply_transforms(subject2, trsfm_list=trsfm_hist, seeds_list=seeds_hist) |
|
84
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
85
|
|
|
self.assertTensorEqual(data1, data2) |
|
86
|
|
|
|
|
87
|
|
|
def test_reproducibility_oneof(self): |
|
88
|
|
|
subject1, subject2 = self.get_subjects() |
|
89
|
|
|
trsfm = Compose([OneOf([RandomNoise(p=1.0), RandomSpike(num_spikes=3, p=1.0)]), RandomNoise(p=.5)]) |
|
90
|
|
|
transformed1 = trsfm(subject1) |
|
91
|
|
|
history1 = transformed1.history |
|
92
|
|
|
trsfm_hist, seeds_hist = compose_from_history(history=history1) |
|
93
|
|
|
transformed2 = self.apply_transforms(subject2, trsfm_list=trsfm_hist, seeds_list=seeds_hist) |
|
94
|
|
|
data1, data2 = transformed1.img.data, transformed2.img.data |
|
95
|
|
|
self.assertTensorEqual(data1, data2) |
|
96
|
|
|
|
|
97
|
|
|
def test_all_random_transforms(self): |
|
98
|
|
|
sample = Subject( |
|
99
|
|
|
t1=ScalarImage(tensor=torch.rand(1, 20, 20, 20)), |
|
100
|
|
|
seg=LabelMap(tensor=torch.rand(1, 20, 20, 20) > 1) |
|
101
|
|
|
) |
|
102
|
|
|
|
|
103
|
|
|
transforms_names = [ |
|
104
|
|
|
name |
|
105
|
|
|
for name in dir(torchio) |
|
106
|
|
|
if name.startswith('Random') |
|
107
|
|
|
] |
|
108
|
|
|
|
|
109
|
|
|
# Downsample at the end so that image shape is not modified |
|
110
|
|
|
transforms_names.remove('RandomDownsample') |
|
111
|
|
|
transforms_names.append('RandomDownsample') |
|
112
|
|
|
|
|
113
|
|
|
transforms = [] |
|
114
|
|
|
for transform_name in transforms_names: |
|
115
|
|
|
# Only transform needing an argument for __init__ |
|
116
|
|
|
if transform_name == 'RandomLabelsToImage': |
|
117
|
|
|
transform = getattr(torchio, transform_name)(label_key='seg') |
|
118
|
|
|
else: |
|
119
|
|
|
transform = getattr(torchio, transform_name)() |
|
120
|
|
|
transforms.append(transform) |
|
121
|
|
|
composed_transform = torchio.Compose(transforms) |
|
122
|
|
|
with warnings.catch_warnings(): # ignore elastic deformation warning |
|
123
|
|
|
warnings.simplefilter('ignore', RuntimeWarning) |
|
124
|
|
|
transformed = composed_transform(sample) |
|
125
|
|
|
|
|
126
|
|
|
new_transforms, seeds = compose_from_history(transformed.history) |
|
127
|
|
|
new_transformed = self.apply_transforms(subject=sample, trsfm_list=new_transforms, seeds_list=seeds) |
|
128
|
|
|
""" |
|
129
|
|
|
new_transforms = [] |
|
130
|
|
|
seeds = [] |
|
131
|
|
|
|
|
132
|
|
|
for transform_name, params_dict in transformed.history: |
|
133
|
|
|
# The Resample transform in the history comes from the DownSampling |
|
134
|
|
|
if transform_name in ['Resample', 'Compose']: |
|
135
|
|
|
continue |
|
136
|
|
|
transform_class = getattr(torchio, transform_name) |
|
137
|
|
|
|
|
138
|
|
|
if transform_name == 'RandomLabelsToImage': |
|
139
|
|
|
transform = transform_class(label_key='seg') |
|
140
|
|
|
else: |
|
141
|
|
|
transform = transform_class() |
|
142
|
|
|
new_transforms.append(transform) |
|
143
|
|
|
seeds.append(params_dict['seed']) |
|
144
|
|
|
|
|
145
|
|
|
composed_transform = torchio.Compose(new_transforms) |
|
146
|
|
|
with warnings.catch_warnings(): # ignore elastic deformation warning |
|
147
|
|
|
warnings.simplefilter('ignore', RuntimeWarning) |
|
148
|
|
|
new_transformed = composed_transform(sample, seeds=seeds) |
|
149
|
|
|
""" |
|
150
|
|
|
|
|
151
|
|
|
self.assertTensorEqual(transformed.t1.data, new_transformed.t1.data) |
|
152
|
|
|
self.assertTensorEqual(transformed.seg.data, new_transformed.seg.data) |
|
153
|
|
|
|
fepegar /
torchio
