|
1
|
|
|
# coding=utf-8 |
|
2
|
|
|
|
|
3
|
|
|
""" |
|
4
|
|
|
Tests for deepreg/model/loss/label.py in |
|
5
|
|
|
pytest style |
|
6
|
|
|
""" |
|
7
|
|
|
|
|
8
|
|
|
from test.unit.util import is_equal_tf |
|
9
|
|
|
|
|
10
|
|
|
import numpy as np |
|
11
|
|
|
import pytest |
|
12
|
|
|
import tensorflow as tf |
|
13
|
|
|
|
|
14
|
|
|
from deepreg.loss.util import MultiScaleMixin, NegativeLossMixin, separable_filter |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
|
|
class TestMultiScaleLoss: |
|
18
|
|
|
def test_get_config(self): |
|
19
|
|
|
loss = MultiScaleMixin() |
|
20
|
|
|
got = loss.get_config() |
|
21
|
|
|
expected = dict( |
|
22
|
|
|
scales=None, |
|
23
|
|
|
kernel="gaussian", |
|
24
|
|
|
reduction=tf.keras.losses.Reduction.AUTO, |
|
25
|
|
|
name="MultiScaleMixin", |
|
26
|
|
|
) |
|
27
|
|
|
assert got == expected |
|
28
|
|
|
|
|
29
|
|
|
|
|
30
|
|
|
def test_separable_filter(): |
|
31
|
|
|
""" |
|
32
|
|
|
Testing separable filter case where non |
|
33
|
|
|
zero length tensor is passed to the |
|
34
|
|
|
function. |
|
35
|
|
|
""" |
|
36
|
|
|
k = np.ones((3, 3, 3, 3, 1), dtype=np.float32) |
|
37
|
|
|
array_eye = np.identity(3, dtype=np.float32) |
|
38
|
|
|
tensor_pred = np.zeros((3, 3, 3, 3, 1), dtype=np.float32) |
|
39
|
|
|
tensor_pred[:, :, 0, 0, 0] = array_eye |
|
40
|
|
|
tensor_pred = tf.convert_to_tensor(tensor_pred, dtype=tf.float32) |
|
41
|
|
|
k = tf.convert_to_tensor(k, dtype=tf.float32) |
|
42
|
|
|
|
|
43
|
|
|
expect = np.ones((3, 3, 3, 3, 1), dtype=np.float32) |
|
44
|
|
|
expect = tf.convert_to_tensor(expect, dtype=tf.float32) |
|
45
|
|
|
|
|
46
|
|
|
get = separable_filter(tensor_pred, k) |
|
47
|
|
|
assert is_equal_tf(get, expect) |
|
48
|
|
|
|
|
49
|
|
|
|
|
50
|
|
|
class MinusClass(tf.keras.losses.Loss): |
|
51
|
|
|
def __init__(self): |
|
52
|
|
|
super().__init__() |
|
53
|
|
|
self.name = "MinusClass" |
|
54
|
|
|
|
|
55
|
|
|
def call(self, y_true, y_pred): |
|
56
|
|
|
return y_true - y_pred |
|
57
|
|
|
|
|
58
|
|
|
|
|
59
|
|
|
class MinusClassLoss(NegativeLossMixin, MinusClass): |
|
60
|
|
|
pass |
|
61
|
|
|
|
|
62
|
|
|
|
|
63
|
|
|
@pytest.mark.parametrize("y_true,y_pred,expected", [(1, 2, 1), (2, 1, -1), (0, 0, 0)]) |
|
64
|
|
|
def test_negative_loss_mixin(y_true, y_pred, expected): |
|
65
|
|
|
""" |
|
66
|
|
|
Testing NegativeLossMixin class that |
|
67
|
|
|
inverts the sign of any value |
|
68
|
|
|
returned by a function |
|
69
|
|
|
|
|
70
|
|
|
:param y_true: int |
|
71
|
|
|
:param y_pred: int |
|
72
|
|
|
:param expected: int |
|
73
|
|
|
:return: |
|
74
|
|
|
""" |
|
75
|
|
|
|
|
76
|
|
|
y_true = tf.constant(y_true, dtype=tf.float32) |
|
77
|
|
|
y_pred = tf.constant(y_pred, dtype=tf.float32) |
|
78
|
|
|
|
|
79
|
|
|
got = MinusClassLoss().call( |
|
80
|
|
|
y_true, |
|
81
|
|
|
y_pred, |
|
82
|
|
|
) |
|
83
|
|
|
assert is_equal_tf(got, expected) |
|
84
|
|
|
|
DeepRegNet /
DeepReg
