1
|
|
|
"""Provide helper functions or classes for defining loss or metrics.""" |
2
|
|
|
|
3
|
|
|
from typing import List, Optional, Union |
4
|
|
|
|
5
|
|
|
import tensorflow as tf |
6
|
|
|
|
7
|
|
|
from deepreg.loss.kernel import cauchy_kernel1d |
8
|
|
|
from deepreg.loss.kernel import gaussian_kernel1d_sigma as gaussian_kernel1d |
9
|
|
|
|
10
|
|
|
|
11
|
|
|
class MultiScaleMixin(tf.keras.losses.Loss): |
12
|
|
|
""" |
13
|
|
|
Mixin class for multi-scale loss. |
14
|
|
|
|
15
|
|
|
It applies the loss at different scales (gaussian or cauchy smoothing). |
16
|
|
|
It is assumed that loss values are between 0 and 1. |
17
|
|
|
""" |
18
|
|
|
|
19
|
|
|
kernel_fn_dict = dict(gaussian=gaussian_kernel1d, cauchy=cauchy_kernel1d) |
20
|
|
|
|
21
|
|
|
def __init__( |
22
|
|
|
self, |
23
|
|
|
scales: Optional[Union[List, float, int]] = None, |
24
|
|
|
kernel: str = "gaussian", |
25
|
|
|
**kwargs, |
26
|
|
|
): |
27
|
|
|
""" |
28
|
|
|
Init. |
29
|
|
|
|
30
|
|
|
:param scales: list of scalars or None, if None, do not apply any scaling. |
31
|
|
|
:param kernel: gaussian or cauchy. |
32
|
|
|
:param kwargs: additional arguments. |
33
|
|
|
""" |
34
|
|
|
super().__init__(**kwargs) |
35
|
|
|
if kernel not in self.kernel_fn_dict: |
36
|
|
|
raise ValueError( |
37
|
|
|
f"Kernel {kernel} is not supported." |
38
|
|
|
f"Supported kernels are {list(self.kernel_fn_dict.keys())}" |
39
|
|
|
) |
40
|
|
|
if scales is not None and not isinstance(scales, list): |
41
|
|
|
scales = [scales] |
42
|
|
|
self.scales = scales |
43
|
|
|
self.kernel = kernel |
44
|
|
|
|
45
|
|
|
def call(self, y_true: tf.Tensor, y_pred: tf.Tensor) -> tf.Tensor: |
46
|
|
|
""" |
47
|
|
|
Use super().call to calculate loss at different scales. |
48
|
|
|
|
49
|
|
|
:param y_true: ground-truth tensor, shape = (batch, dim1, dim2, dim3). |
50
|
|
|
:param y_pred: predicted tensor, shape = (batch, dim1, dim2, dim3). |
51
|
|
|
:return: multi-scale loss, shape = (batch, ). |
52
|
|
|
""" |
53
|
|
|
if self.scales is None: |
54
|
|
|
return super().call(y_true=y_true, y_pred=y_pred) |
55
|
|
|
kernel_fn = self.kernel_fn_dict[self.kernel] |
56
|
|
|
losses = [] |
57
|
|
|
for s in self.scales: |
58
|
|
|
if s == 0: |
59
|
|
|
# no smoothing |
60
|
|
|
losses.append( |
61
|
|
|
super().call( |
62
|
|
|
y_true=y_true, |
63
|
|
|
y_pred=y_pred, |
64
|
|
|
) |
65
|
|
|
) |
66
|
|
|
else: |
67
|
|
|
losses.append( |
68
|
|
|
super().call( |
69
|
|
|
y_true=separable_filter( |
70
|
|
|
tf.expand_dims(y_true, axis=4), kernel_fn(s) |
71
|
|
|
)[..., 0], |
72
|
|
|
y_pred=separable_filter( |
73
|
|
|
tf.expand_dims(y_pred, axis=4), kernel_fn(s) |
74
|
|
|
)[..., 0], |
75
|
|
|
) |
76
|
|
|
) |
77
|
|
|
loss = tf.add_n(losses) |
78
|
|
|
loss = loss / len(self.scales) |
79
|
|
|
return loss |
80
|
|
|
|
81
|
|
|
def get_config(self) -> dict: |
82
|
|
|
"""Return the config dictionary for recreating this class.""" |
83
|
|
|
config = super().get_config() |
84
|
|
|
config["scales"] = self.scales |
85
|
|
|
config["kernel"] = self.kernel |
86
|
|
|
return config |
87
|
|
|
|
88
|
|
|
|
89
|
|
|
class NegativeLossMixin(tf.keras.losses.Loss): |
90
|
|
|
"""Mixin class to revert the sign of the loss value.""" |
91
|
|
|
|
92
|
|
|
def __init__(self, **kwargs): |
93
|
|
|
""" |
94
|
|
|
Init without required arguments. |
95
|
|
|
|
96
|
|
|
:param kwargs: additional arguments. |
97
|
|
|
""" |
98
|
|
|
super().__init__(**kwargs) |
99
|
|
|
self.name = self.name + "Loss" |
100
|
|
|
|
101
|
|
|
def call(self, y_true: tf.Tensor, y_pred: tf.Tensor) -> tf.Tensor: |
102
|
|
|
""" |
103
|
|
|
Revert the sign of loss. |
104
|
|
|
|
105
|
|
|
:param y_true: ground-truth tensor. |
106
|
|
|
:param y_pred: predicted tensor. |
107
|
|
|
:return: negated loss. |
108
|
|
|
""" |
109
|
|
|
return -super().call(y_true=y_true, y_pred=y_pred) |
110
|
|
|
|
111
|
|
|
|
112
|
|
|
def separable_filter(tensor: tf.Tensor, kernel: tf.Tensor) -> tf.Tensor: |
113
|
|
|
""" |
114
|
|
|
Create a 3d separable filter. |
115
|
|
|
|
116
|
|
|
Here `tf.nn.conv3d` accepts the `filters` argument of shape |
117
|
|
|
(filter_depth, filter_height, filter_width, in_channels, out_channels), |
118
|
|
|
where the first axis of `filters` is the depth not batch, |
119
|
|
|
and the input to `tf.nn.conv3d` is of shape |
120
|
|
|
(batch, in_depth, in_height, in_width, in_channels). |
121
|
|
|
|
122
|
|
|
:param tensor: shape = (batch, dim1, dim2, dim3, 1) |
123
|
|
|
:param kernel: shape = (dim4,) |
124
|
|
|
:return: shape = (batch, dim1, dim2, dim3, 1) |
125
|
|
|
""" |
126
|
|
|
strides = [1, 1, 1, 1, 1] |
127
|
|
|
kernel = tf.cast(kernel, dtype=tensor.dtype) |
128
|
|
|
|
129
|
|
|
tensor = tf.nn.conv3d( |
130
|
|
|
tf.nn.conv3d( |
131
|
|
|
tf.nn.conv3d( |
132
|
|
|
tensor, |
133
|
|
|
filters=tf.reshape(kernel, [-1, 1, 1, 1, 1]), |
134
|
|
|
strides=strides, |
135
|
|
|
padding="SAME", |
136
|
|
|
), |
137
|
|
|
filters=tf.reshape(kernel, [1, -1, 1, 1, 1]), |
138
|
|
|
strides=strides, |
139
|
|
|
padding="SAME", |
140
|
|
|
), |
141
|
|
|
filters=tf.reshape(kernel, [1, 1, -1, 1, 1]), |
142
|
|
|
strides=strides, |
143
|
|
|
padding="SAME", |
144
|
|
|
) |
145
|
|
|
return tensor |
146
|
|
|
|