1
|
|
|
# coding=utf-8 |
2
|
|
|
|
3
|
|
|
""" |
4
|
|
|
Tests for deepreg/model/layer |
5
|
|
|
""" |
6
|
|
|
from typing import Optional |
7
|
|
|
|
8
|
|
|
import numpy as np |
9
|
|
|
import pytest |
10
|
|
|
import tensorflow as tf |
11
|
|
|
|
12
|
|
|
import deepreg.model.layer as layer |
13
|
|
|
|
14
|
|
|
|
15
|
|
|
@pytest.mark.parametrize( |
16
|
|
|
("input_shape", "output_shape", "expected_shape"), |
17
|
|
|
[ |
18
|
|
|
((6, 7, 8), (12, 14, 16), (12, 14, 16)), |
19
|
|
|
((6, 7, 8), (11, 13, 15), (11, 13, 15)), |
20
|
|
|
((6, 7, 8), None, (12, 14, 16)), |
21
|
|
|
], |
|
|
|
|
22
|
|
|
) |
23
|
|
|
def test_deconv3d(input_shape, output_shape: Optional[tuple], expected_shape: tuple): |
24
|
|
|
""" |
25
|
|
|
Test output shapes and configs. |
26
|
|
|
|
27
|
|
|
:param input_shape: input shape for layer definition |
28
|
|
|
:param output_shape: output shape for layer definition |
29
|
|
|
:param expected_shape: expected output shape |
30
|
|
|
""" |
31
|
|
|
batch_size = 5 |
32
|
|
|
|
33
|
|
|
deconv3d = layer.Deconv3d(filters=3, strides=2, output_shape=output_shape) |
34
|
|
|
|
35
|
|
|
inputs = tf.ones(shape=(batch_size,) + input_shape + (1,)) |
36
|
|
|
output = deconv3d(inputs) |
37
|
|
|
|
38
|
|
|
assert output.shape == (batch_size,) + expected_shape + (3,) |
39
|
|
|
|
40
|
|
|
config = deconv3d.get_config() |
41
|
|
|
assert config == dict( |
42
|
|
|
filters=3, |
43
|
|
|
output_shape=output_shape, |
44
|
|
|
kernel_size=3, |
45
|
|
|
strides=2, |
46
|
|
|
padding="same", |
47
|
|
|
name="deconv3d", |
48
|
|
|
trainable=True, |
49
|
|
|
dtype="float32", |
50
|
|
|
) |
51
|
|
|
|
52
|
|
|
|
53
|
|
|
@pytest.mark.parametrize("layer_name", ["conv3d", "deconv3d"]) |
54
|
|
|
@pytest.mark.parametrize("norm_name", ["batch", "layer"]) |
55
|
|
|
@pytest.mark.parametrize("activation", ["relu", "elu"]) |
56
|
|
|
def test_norm_block(layer_name: str, norm_name: str, activation: str): |
57
|
|
|
""" |
58
|
|
|
Test output shapes and configs. |
59
|
|
|
|
60
|
|
|
:param layer_name: layer_name for layer definition |
61
|
|
|
:param norm_name: norm_name for layer definition |
62
|
|
|
:param activation: activation for layer definition |
63
|
|
|
""" |
64
|
|
|
input_size = (2, 3, 4, 5, 6) # (batch, *shape, ch) |
65
|
|
|
conv_block = layer.NormBlock( |
66
|
|
|
layer_name=layer_name, |
67
|
|
|
norm_name=norm_name, |
68
|
|
|
activation=activation, |
69
|
|
|
filters=3, |
70
|
|
|
kernel_size=1, |
71
|
|
|
padding="same", |
72
|
|
|
) |
73
|
|
|
inputs = tf.ones(shape=input_size) |
74
|
|
|
outputs = conv_block(inputs) |
75
|
|
|
assert outputs.shape == input_size[:-1] + (3,) |
76
|
|
|
|
77
|
|
|
config = conv_block.get_config() |
78
|
|
|
assert config == dict( |
79
|
|
|
layer_name=layer_name, |
80
|
|
|
norm_name=norm_name, |
81
|
|
|
activation=activation, |
82
|
|
|
filters=3, |
83
|
|
|
kernel_size=1, |
84
|
|
|
padding="same", |
85
|
|
|
name="norm_block", |
86
|
|
|
trainable=True, |
87
|
|
|
dtype="float32", |
88
|
|
|
) |
89
|
|
|
|
90
|
|
|
|
91
|
|
|
def test_residual3d_block(): |
92
|
|
|
""" |
93
|
|
|
Test the layer.Residual3dBlock class and its default attributes. |
94
|
|
|
""" |
95
|
|
|
res3d_block = layer.Residual3dBlock(8) |
96
|
|
|
|
97
|
|
|
assert isinstance(res3d_block._conv3d_block, layer.Conv3dBlock) |
|
|
|
|
98
|
|
|
assert res3d_block._conv3d_block._conv3d.kernel_size == (3, 3, 3) |
|
|
|
|
99
|
|
|
assert res3d_block._conv3d_block._conv3d.strides == (1, 1, 1) |
|
|
|
|
100
|
|
|
|
101
|
|
|
assert res3d_block._conv3d.use_bias is False |
|
|
|
|
102
|
|
|
assert res3d_block._conv3d.kernel_size == (3, 3, 3) |
|
|
|
|
103
|
|
|
assert res3d_block._conv3d.strides == (1, 1, 1) |
|
|
|
|
104
|
|
|
|
105
|
|
|
|
106
|
|
|
def test_downsample_resnet_block(): |
107
|
|
|
""" |
108
|
|
|
Test the layer.DownSampleResnetBlock class and its default attributes. |
109
|
|
|
""" |
110
|
|
|
model = layer.DownSampleResnetBlock(8) |
111
|
|
|
|
112
|
|
|
assert model._pooling is True |
|
|
|
|
113
|
|
|
|
114
|
|
|
assert isinstance(model._residual_block, layer.Residual3dBlock) |
|
|
|
|
115
|
|
|
assert model._conv3d_block3 is None |
|
|
|
|
116
|
|
|
|
117
|
|
|
model = layer.DownSampleResnetBlock(8, pooling=False) |
118
|
|
|
assert model._max_pool3d is None |
|
|
|
|
119
|
|
|
assert isinstance(model._conv3d_block3, layer.Conv3dBlock) |
|
|
|
|
120
|
|
|
|
121
|
|
|
|
122
|
|
|
def test_upsample_resnet_block(): |
123
|
|
|
""" |
124
|
|
|
Test the layer.UpSampleResnetBlock class and its default attributes. |
125
|
|
|
""" |
126
|
|
|
batch_size = 5 |
127
|
|
|
channels = 4 |
128
|
|
|
input_size = (32, 32, 16) |
129
|
|
|
output_size = (64, 64, 32) |
130
|
|
|
|
131
|
|
|
input_tensor_size = (batch_size,) + input_size + (channels,) |
132
|
|
|
skip_tensor_size = (batch_size,) + output_size + (channels // 2,) |
133
|
|
|
|
134
|
|
|
model = layer.UpSampleResnetBlock(8) |
135
|
|
|
model.build([input_tensor_size, skip_tensor_size]) |
136
|
|
|
|
137
|
|
|
assert model._filters == 8 |
|
|
|
|
138
|
|
|
assert model._concat is False |
|
|
|
|
139
|
|
|
assert isinstance(model._conv3d_block, layer.Conv3dBlock) |
|
|
|
|
140
|
|
|
assert isinstance(model._residual_block, layer.Residual3dBlock) |
|
|
|
|
141
|
|
|
assert isinstance(model._deconv3d_block, layer.Deconv3dBlock) |
|
|
|
|
142
|
|
|
|
143
|
|
|
|
144
|
|
|
def test_init_conv3d_with_resize(): |
145
|
|
|
""" |
146
|
|
|
Test the layer.Conv3dWithResize class's default attributes and call function. |
147
|
|
|
""" |
148
|
|
|
batch_size = 5 |
149
|
|
|
channels = 4 |
150
|
|
|
input_size = (32, 32, 16) |
151
|
|
|
output_size = (62, 62, 24) |
152
|
|
|
filters = 8 |
153
|
|
|
|
154
|
|
|
input_tensor_size = (batch_size,) + input_size + (channels,) |
155
|
|
|
output_tensor_size = (batch_size,) + output_size + (filters,) |
156
|
|
|
|
157
|
|
|
model = layer.Conv3dWithResize(output_size, filters) |
158
|
|
|
|
159
|
|
|
assert model._output_shape == output_size |
|
|
|
|
160
|
|
|
|
161
|
|
|
# Pass an input of all zeros |
162
|
|
|
inputs = np.zeros(input_tensor_size) |
163
|
|
|
# Get outputs by calling |
164
|
|
|
output = model.call(inputs) |
165
|
|
|
# Expected shape is (5, 1, 2, 3, 3) |
166
|
|
|
assert all(x == y for x, y in zip(output_tensor_size, output.shape)) |
167
|
|
|
|
168
|
|
|
|
169
|
|
|
def test_warping(): |
170
|
|
|
""" |
171
|
|
|
Test the layer.Warping class, its default attributes and its call() method. |
172
|
|
|
""" |
173
|
|
|
batch_size = 5 |
174
|
|
|
fixed_image_size = (32, 32, 16) |
175
|
|
|
moving_image_size = (24, 24, 16) |
176
|
|
|
ndims = len(moving_image_size) |
177
|
|
|
|
178
|
|
|
grid_size = (1,) + fixed_image_size + (3,) |
179
|
|
|
model = layer.Warping(fixed_image_size) |
180
|
|
|
|
181
|
|
|
assert all(x == y for x, y in zip(grid_size, model.grid_ref.shape)) |
182
|
|
|
|
183
|
|
|
# Pass an input of all zeros |
184
|
|
|
inputs = [ |
185
|
|
|
np.ones((batch_size, *fixed_image_size, ndims), dtype="float32"), |
186
|
|
|
np.ones((batch_size, *moving_image_size), dtype="float32"), |
187
|
|
|
] |
188
|
|
|
# Get outputs by calling |
189
|
|
|
output = model.call(inputs) |
190
|
|
|
# Expected shape is (5, 1, 2, 3, 3) |
191
|
|
|
assert all(x == y for x, y in zip((batch_size,) + fixed_image_size, output.shape)) |
192
|
|
|
|
193
|
|
|
|
194
|
|
|
def test_init_dvf(): |
195
|
|
|
""" |
196
|
|
|
Test the layer.IntDVF class, its default attributes and its call() method. |
197
|
|
|
""" |
198
|
|
|
|
199
|
|
|
batch_size = 5 |
200
|
|
|
fixed_image_size = (32, 32, 16) |
201
|
|
|
ndims = len(fixed_image_size) |
202
|
|
|
|
203
|
|
|
model = layer.IntDVF(fixed_image_size) |
204
|
|
|
|
205
|
|
|
assert isinstance(model._warping, layer.Warping) |
|
|
|
|
206
|
|
|
assert model._num_steps == 7 |
|
|
|
|
207
|
|
|
|
208
|
|
|
inputs = np.ones((batch_size, *fixed_image_size, ndims)) |
209
|
|
|
output = model.call(inputs) |
210
|
|
|
assert all( |
211
|
|
|
x == y |
212
|
|
|
for x, y in zip((batch_size,) + fixed_image_size + (ndims,), output.shape) |
213
|
|
|
) |
214
|
|
|
|
215
|
|
|
|
216
|
|
|
def test_additive_upsampling(): |
217
|
|
|
""" |
218
|
|
|
Test the layer.AdditiveUpSampling class and its default attributes. |
219
|
|
|
""" |
220
|
|
|
channels = 8 |
221
|
|
|
batch_size = 5 |
222
|
|
|
output_size = (32, 32, 16) |
223
|
|
|
input_size = (24, 24, 16) |
224
|
|
|
|
225
|
|
|
# Test __init__() |
226
|
|
|
model = layer.AdditiveUpSampling(output_size) |
227
|
|
|
assert model._stride == 2 |
|
|
|
|
228
|
|
|
assert model._output_shape == output_size |
|
|
|
|
229
|
|
|
|
230
|
|
|
# Test call() |
231
|
|
|
inputs = np.ones( |
232
|
|
|
(batch_size, input_size[0], input_size[1], input_size[2], channels) |
233
|
|
|
) |
234
|
|
|
output = model(inputs) |
235
|
|
|
assert all( |
236
|
|
|
x == y |
237
|
|
|
for x, y in zip((batch_size,) + output_size + (channels / 2,), output.shape) |
238
|
|
|
) |
239
|
|
|
|
240
|
|
|
# Test the exceptions |
241
|
|
|
model = layer.AdditiveUpSampling(output_size, stride=3) |
242
|
|
|
with pytest.raises(ValueError): |
243
|
|
|
model(inputs) |
244
|
|
|
|
245
|
|
|
|
246
|
|
|
def test_local_net_residual3d_block(): |
247
|
|
|
""" |
248
|
|
|
Test the layer.LocalNetResidual3dBlock class's, |
249
|
|
|
default attributes and call() function. |
250
|
|
|
""" |
251
|
|
|
|
252
|
|
|
# Test __init__() |
253
|
|
|
conv3d_block = layer.LocalNetResidual3dBlock(8) |
254
|
|
|
|
255
|
|
|
assert conv3d_block._conv3d.kernel_size == (3, 3, 3) |
|
|
|
|
256
|
|
|
assert conv3d_block._conv3d.strides == (1, 1, 1) |
|
|
|
|
257
|
|
|
assert conv3d_block._conv3d.padding == "same" |
|
|
|
|
258
|
|
|
assert conv3d_block._conv3d.use_bias is False |
|
|
|
|
259
|
|
|
|
260
|
|
|
|
261
|
|
|
def test_local_net_upsample_resnet_block(): |
262
|
|
|
""" |
263
|
|
|
Test the layer.LocalNetUpSampleResnetBlock class, |
264
|
|
|
its default attributes and its call() function. |
265
|
|
|
""" |
266
|
|
|
batch_size = 5 |
267
|
|
|
channels = 4 |
268
|
|
|
input_size = (32, 32, 16) |
269
|
|
|
output_size = (64, 64, 32) |
270
|
|
|
|
271
|
|
|
nonskip_tensor_size = (batch_size,) + input_size + (channels,) |
272
|
|
|
skip_tensor_size = (batch_size,) + output_size + (channels,) |
273
|
|
|
|
274
|
|
|
# Test __init__() and build() |
275
|
|
|
model = layer.LocalNetUpSampleResnetBlock(8) |
276
|
|
|
model.build([nonskip_tensor_size, skip_tensor_size]) |
277
|
|
|
|
278
|
|
|
assert model._filters == 8 |
|
|
|
|
279
|
|
|
assert model._use_additive_upsampling is True |
|
|
|
|
280
|
|
|
|
281
|
|
|
assert isinstance(model._deconv3d_block, layer.Deconv3dBlock) |
|
|
|
|
282
|
|
|
assert isinstance(model._additive_upsampling, layer.AdditiveUpSampling) |
|
|
|
|
283
|
|
|
assert isinstance(model._conv3d_block, layer.Conv3dBlock) |
|
|
|
|
284
|
|
|
assert isinstance(model._residual_block, layer.LocalNetResidual3dBlock) |
|
|
|
|
285
|
|
|
|
286
|
|
|
|
287
|
|
|
class TestResizeCPTransform: |
|
|
|
|
288
|
|
|
@pytest.mark.parametrize( |
289
|
|
|
"parameter,cp_spacing", [((8, 8, 8), 8), ((8, 24, 16), (8, 24, 16))] |
290
|
|
|
) |
|
|
|
|
291
|
|
|
def test_attributes(self, parameter, cp_spacing): |
292
|
|
|
model = layer.ResizeCPTransform(cp_spacing) |
293
|
|
|
|
294
|
|
|
if isinstance(cp_spacing, int): |
295
|
|
|
cp_spacing = [cp_spacing] * 3 |
296
|
|
|
assert list(model.cp_spacing) == list(parameter) |
297
|
|
|
assert model.kernel_sigma == [0.44 * cp for cp in cp_spacing] |
298
|
|
|
|
299
|
|
|
@pytest.mark.parametrize( |
300
|
|
|
"input_size,output_size,cp_spacing", |
301
|
|
|
[ |
302
|
|
|
((1, 8, 8, 8, 3), (12, 8, 12), (8, 16, 8)), |
303
|
|
|
((1, 8, 8, 8, 3), (12, 12, 12), 8), |
304
|
|
|
], |
|
|
|
|
305
|
|
|
) |
306
|
|
|
def test_build(self, input_size, output_size, cp_spacing): |
307
|
|
|
model = layer.ResizeCPTransform(cp_spacing) |
308
|
|
|
model.build(input_size) |
309
|
|
|
|
310
|
|
|
assert [a == b for a, b, in zip(model._output_shape, output_size)] |
|
|
|
|
311
|
|
|
|
312
|
|
|
@pytest.mark.parametrize( |
313
|
|
|
"input_size,output_size,cp_spacing", |
314
|
|
|
[ |
315
|
|
|
((1, 68, 68, 68, 3), (1, 12, 8, 12, 3), (8, 16, 8)), |
316
|
|
|
((1, 68, 68, 68, 3), (1, 12, 12, 12, 3), 8), |
317
|
|
|
], |
|
|
|
|
318
|
|
|
) |
319
|
|
|
def test_call(self, input_size, output_size, cp_spacing): |
320
|
|
|
model = layer.ResizeCPTransform(cp_spacing) |
321
|
|
|
model.build(input_size) |
322
|
|
|
|
323
|
|
|
input = tf.random.normal(shape=input_size, dtype=tf.float32) |
|
|
|
|
324
|
|
|
output = model(input) |
325
|
|
|
|
326
|
|
|
assert output.shape == output_size |
327
|
|
|
|
328
|
|
|
|
329
|
|
|
class TestBSplines3DTransform: |
330
|
|
|
""" |
331
|
|
|
Test the layer.BSplines3DTransform class, |
332
|
|
|
its default attributes and its call() function. |
333
|
|
|
""" |
334
|
|
|
|
335
|
|
|
@pytest.mark.parametrize( |
336
|
|
|
"input_size,cp", |
337
|
|
|
[((1, 8, 8, 8, 3), 8), ((1, 8, 8, 8, 3), (8, 16, 12))], |
338
|
|
|
) |
|
|
|
|
339
|
|
|
def test_init(self, input_size, cp): |
340
|
|
|
model = layer.BSplines3DTransform(cp, input_size[1:-1]) |
341
|
|
|
|
342
|
|
|
if isinstance(cp, int): |
343
|
|
|
cp = (cp, cp, cp) |
344
|
|
|
|
345
|
|
|
assert model.cp_spacing == cp |
346
|
|
|
|
347
|
|
|
@pytest.mark.parametrize( |
348
|
|
|
"input_size,cp", |
349
|
|
|
[((1, 8, 8, 8, 3), (8, 8, 8)), ((1, 8, 8, 8, 3), (8, 16, 12))], |
350
|
|
|
) |
|
|
|
|
351
|
|
|
def generate_filter_coefficients(self, cp_spacing): |
352
|
|
|
|
353
|
|
|
b = { |
354
|
|
|
0: lambda u: np.float64((1 - u) ** 3 / 6), |
355
|
|
|
1: lambda u: np.float64((3 * (u ** 3) - 6 * (u ** 2) + 4) / 6), |
356
|
|
|
2: lambda u: np.float64((-3 * (u ** 3) + 3 * (u ** 2) + 3 * u + 1) / 6), |
357
|
|
|
3: lambda u: np.float64(u ** 3 / 6), |
358
|
|
|
} |
359
|
|
|
|
360
|
|
|
filters = np.zeros( |
361
|
|
|
( |
362
|
|
|
4 * cp_spacing[0], |
363
|
|
|
4 * cp_spacing[1], |
364
|
|
|
4 * cp_spacing[2], |
365
|
|
|
3, |
366
|
|
|
3, |
367
|
|
|
), |
368
|
|
|
dtype=np.float32, |
369
|
|
|
) |
370
|
|
|
|
371
|
|
|
for u in range(cp_spacing[0]): |
|
|
|
|
372
|
|
|
for v in range(cp_spacing[1]): |
373
|
|
|
for w in range(cp_spacing[2]): |
374
|
|
|
for x in range(4): |
375
|
|
|
for y in range(4): |
376
|
|
|
for z in range(4): |
377
|
|
|
for it_dim in range(3): |
378
|
|
|
u_norm = 1 - (u + 0.5) / cp_spacing[0] |
379
|
|
|
v_norm = 1 - (v + 0.5) / cp_spacing[1] |
380
|
|
|
w_norm = 1 - (w + 0.5) / cp_spacing[2] |
381
|
|
|
filters[ |
382
|
|
|
x * cp_spacing[0] + u, |
383
|
|
|
y * cp_spacing[1] + v, |
384
|
|
|
z * cp_spacing[2] + w, |
385
|
|
|
it_dim, |
386
|
|
|
it_dim, |
387
|
|
|
] = ( |
388
|
|
|
b[x](u_norm) * b[y](v_norm) * b[z](w_norm) |
389
|
|
|
) |
390
|
|
|
return filters |
391
|
|
|
|
392
|
|
|
@pytest.mark.parametrize( |
393
|
|
|
"input_size,cp", |
394
|
|
|
[((1, 8, 8, 8, 3), (8, 8, 8)), ((1, 8, 8, 8, 3), (8, 16, 12))], |
395
|
|
|
) |
|
|
|
|
396
|
|
|
def test_build(self, input_size, cp): |
397
|
|
|
model = layer.BSplines3DTransform(cp, input_size[1:-1]) |
398
|
|
|
|
399
|
|
|
model.build(input_size) |
400
|
|
|
assert model.filter.shape == ( |
401
|
|
|
4 * cp[0], |
402
|
|
|
4 * cp[1], |
403
|
|
|
4 * cp[2], |
404
|
|
|
3, |
405
|
|
|
3, |
406
|
|
|
) |
407
|
|
|
|
408
|
|
|
@pytest.mark.parametrize( |
409
|
|
|
"input_size,cp", |
410
|
|
|
[((1, 8, 8, 8, 3), (8, 8, 8)), ((1, 8, 8, 8, 3), (8, 16, 12))], |
411
|
|
|
) |
|
|
|
|
412
|
|
|
def test_coefficients(self, input_size, cp): |
413
|
|
|
|
414
|
|
|
filters = self.generate_filter_coefficients(cp_spacing=cp) |
415
|
|
|
|
416
|
|
|
model = layer.BSplines3DTransform(cp, input_size[1:-1]) |
417
|
|
|
model.build(input_size) |
418
|
|
|
|
419
|
|
|
assert np.allclose(filters, model.filter.numpy(), atol=1e-8) |
420
|
|
|
|
421
|
|
|
@pytest.mark.parametrize( |
422
|
|
|
"input_size,cp", |
423
|
|
|
[((1, 8, 8, 8, 3), (8, 8, 8)), ((1, 8, 8, 8, 3), (8, 16, 12))], |
424
|
|
|
) |
|
|
|
|
425
|
|
|
def test_interpolation(self, input_size, cp): |
426
|
|
|
model = layer.BSplines3DTransform(cp, input_size[1:-1]) |
427
|
|
|
model.build(input_size) |
428
|
|
|
|
429
|
|
|
vol_shape = input_size[1:-1] |
430
|
|
|
num_cp = ( |
431
|
|
|
[input_size[0]] |
432
|
|
|
+ [int(np.ceil(isize / cpsize) + 3) for isize, cpsize in zip(vol_shape, cp)] |
433
|
|
|
+ [input_size[-1]] |
434
|
|
|
) |
435
|
|
|
|
436
|
|
|
field = tf.random.normal(shape=num_cp, dtype=tf.float32) |
437
|
|
|
|
438
|
|
|
ddf = model.call(field) |
439
|
|
|
assert ddf.shape == input_size |
440
|
|
|
|