|
1
|
|
|
# coding=utf-8 |
|
2
|
|
|
|
|
3
|
|
|
""" |
|
4
|
|
|
Tests for deepreg/model/backbone/u_net.py |
|
5
|
|
|
""" |
|
6
|
|
|
from typing import Tuple |
|
7
|
|
|
|
|
8
|
|
|
import pytest |
|
9
|
|
|
import tensorflow as tf |
|
10
|
|
|
|
|
11
|
|
|
from deepreg.model.backbone.u_net import UNet |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class TestUNet: |
|
15
|
|
|
""" |
|
16
|
|
|
Test the backbone.u_net.UNet class |
|
17
|
|
|
""" |
|
18
|
|
|
|
|
19
|
|
|
@pytest.mark.parametrize( |
|
20
|
|
|
"depth,encode_num_channels,decode_num_channels", |
|
21
|
|
|
[ |
|
22
|
|
|
(2, (4, 8, 16), (4, 8, 16)), |
|
23
|
|
|
(2, (4, 8, 8), (4, 8, 8)), |
|
24
|
|
|
(2, (4, 8, 8), (8, 8, 8)), |
|
25
|
|
|
], |
|
26
|
|
|
) |
|
27
|
|
|
@pytest.mark.parametrize("pooling", [True, False]) |
|
|
|
|
|
|
28
|
|
|
@pytest.mark.parametrize("concat_skip", [True, False]) |
|
29
|
|
|
def test_channels( |
|
30
|
|
|
self, |
|
31
|
|
|
depth: int, |
|
32
|
|
|
encode_num_channels: Tuple, |
|
33
|
|
|
decode_num_channels: Tuple, |
|
34
|
|
|
pooling: bool, |
|
35
|
|
|
concat_skip: bool, |
|
36
|
|
|
): |
|
37
|
|
|
""" |
|
38
|
|
|
Test unet with custom encode/decode channels. |
|
39
|
|
|
|
|
40
|
|
|
:param depth: input is at level 0, bottom is at level depth |
|
41
|
|
|
:param encode_num_channels: filters/channels for down-sampling, |
|
42
|
|
|
by default it is doubled at each layer during down-sampling |
|
43
|
|
|
:param decode_num_channels: filters/channels for up-sampling, |
|
44
|
|
|
by default it is the same as encode_num_channels |
|
45
|
|
|
:param pooling: for down-sampling, use non-parameterized |
|
46
|
|
|
pooling if true, otherwise use conv3d |
|
47
|
|
|
:param concat_skip: if concatenate skip or add it |
|
48
|
|
|
""" |
|
49
|
|
|
# in case of adding skip tensors, the channels should match |
|
50
|
|
|
expect_err = (not concat_skip) and encode_num_channels != decode_num_channels |
|
51
|
|
|
|
|
52
|
|
|
image_size = (5, 6, 7) |
|
53
|
|
|
out_ch = 3 |
|
54
|
|
|
try: |
|
55
|
|
|
network = UNet( |
|
56
|
|
|
image_size=image_size, |
|
57
|
|
|
out_channels=out_ch, |
|
58
|
|
|
num_channel_initial=None, |
|
59
|
|
|
encode_num_channels=encode_num_channels, |
|
60
|
|
|
decode_num_channels=decode_num_channels, |
|
61
|
|
|
depth=depth, |
|
62
|
|
|
out_kernel_initializer="he_normal", |
|
63
|
|
|
out_activation="softmax", |
|
64
|
|
|
pooling=pooling, |
|
65
|
|
|
concat_skip=concat_skip, |
|
66
|
|
|
) |
|
67
|
|
|
except ValueError as err: |
|
68
|
|
|
if expect_err: |
|
69
|
|
|
return |
|
70
|
|
|
raise err |
|
71
|
|
|
inputs = tf.ones(shape=(5, *image_size, out_ch)) |
|
72
|
|
|
|
|
73
|
|
|
output = network.call(inputs) |
|
74
|
|
|
assert inputs.shape == output.shape |
|
75
|
|
|
|
|
76
|
|
|
@pytest.mark.parametrize( |
|
77
|
|
|
"image_size,depth", |
|
78
|
|
|
[((11, 12, 13), 5), ((8, 8, 8), 3)], |
|
79
|
|
|
) |
|
80
|
|
|
@pytest.mark.parametrize("pooling", [True, False]) |
|
81
|
|
|
@pytest.mark.parametrize("concat_skip", [True, False]) |
|
82
|
|
|
def test_call( |
|
83
|
|
|
self, |
|
84
|
|
|
image_size: Tuple, |
|
85
|
|
|
depth: int, |
|
86
|
|
|
pooling: bool, |
|
87
|
|
|
concat_skip: bool, |
|
88
|
|
|
): |
|
89
|
|
|
""" |
|
90
|
|
|
Test unet call function. |
|
91
|
|
|
|
|
92
|
|
|
:param image_size: (dim1, dim2, dim3), dims of input image. |
|
93
|
|
|
:param depth: input is at level 0, bottom is at level depth |
|
94
|
|
|
:param pooling: for down-sampling, use non-parameterized |
|
95
|
|
|
pooling if true, otherwise use conv3d |
|
96
|
|
|
:param concat_skip: if concatenate skip or add it |
|
97
|
|
|
""" |
|
98
|
|
|
out_ch = 3 |
|
99
|
|
|
network = UNet( |
|
100
|
|
|
image_size=image_size, |
|
101
|
|
|
out_channels=out_ch, |
|
102
|
|
|
num_channel_initial=2, |
|
103
|
|
|
depth=depth, |
|
104
|
|
|
out_kernel_initializer="he_normal", |
|
105
|
|
|
out_activation="softmax", |
|
106
|
|
|
pooling=pooling, |
|
107
|
|
|
concat_skip=concat_skip, |
|
108
|
|
|
) |
|
109
|
|
|
inputs = tf.ones(shape=(5, *image_size, out_ch)) |
|
110
|
|
|
output = network.call(inputs) |
|
111
|
|
|
assert inputs.shape == output.shape |
|
112
|
|
|
|
|
113
|
|
View Code Duplication |
def test_get_config(self): |
|
|
|
|
|
|
114
|
|
|
config = dict( |
|
115
|
|
|
image_size=(4, 5, 6), |
|
116
|
|
|
out_channels=3, |
|
117
|
|
|
num_channel_initial=2, |
|
118
|
|
|
depth=2, |
|
119
|
|
|
extract_levels=(0, 1), |
|
120
|
|
|
out_kernel_initializer="he_normal", |
|
121
|
|
|
out_activation="softmax", |
|
122
|
|
|
pooling=False, |
|
123
|
|
|
concat_skip=False, |
|
124
|
|
|
encode_kernel_sizes=3, |
|
125
|
|
|
decode_kernel_sizes=3, |
|
126
|
|
|
encode_num_channels=(2, 4, 8), |
|
127
|
|
|
decode_num_channels=(2, 4, 8), |
|
128
|
|
|
strides=2, |
|
129
|
|
|
padding="same", |
|
130
|
|
|
name="Test", |
|
131
|
|
|
) |
|
132
|
|
|
network = UNet(**config) |
|
133
|
|
|
got = network.get_config() |
|
134
|
|
|
assert got == config |
|
135
|
|
|
|
DeepRegNet /
DeepReg
Loading history...