-
Notifications
You must be signed in to change notification settings - Fork 0
/
unused_loss_functions.py
224 lines (166 loc) · 9.01 KB
/
unused_loss_functions.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
class WeightedIoULoss(k.losses.Loss):
def __init__(self, class_weights, epsilon=1e-7, name='weighted_iou_loss'):
super(WeightedIoULoss, self).__init__(name=name)
self.class_weights = class_weights
self.epsilon = epsilon
def call(self, y_true, y_pred):
intersection = tf.reduce_sum(y_true * y_pred, axis=[1, 2])
union = tf.reduce_sum(y_true + y_pred, axis=[1, 2]) - intersection
iou = 1 - (intersection + self.epsilon) / (union + self.epsilon)
weighted_iou = iou * self.class_weights
# false_positive_percentage = tf.reduce_mean(tf.clip_by_value(y_pred - y_true, 0, 1), axis=[1,2])
per_class = weighted_iou # + false_positive_percentage
# k.backend.print_tensor(weighted_iou, message='weighted_iou = ')
iou_loss = tf.reduce_mean(per_class)
# k.backend.print_tensor(loss, message='loss first red = ')
# iou_loss = tf.reduce_mean(iou_loss)
# k.backend.print_tensor(loss, message='loss = ')
loss = iou_loss
return loss
class SmoothDiceLoss(k.losses.Loss):
def __init__(self, name="smooth_dice_loss"):
super(SmoothDiceLoss, self).__init__(name=name)
def call(self, y_true, y_pred):
return 1 - dice_coef(y_true, y_pred, self.smooth)
class SmoothDiceLossSampleWise(k.losses.Loss):
def __init__(self, name="smooth_dice_loss_sample_wise"):
super(SmoothDiceLossSampleWise, self).__init__(name=name)
def call(self, y_true, y_pred):
per_sample_loss = tf.map_fn(lambda x: 1 - dice_coef(x[0], x[1]), (y_true, y_pred), dtype=tf.float32)
loss = tf.math.reduce_mean(per_sample_loss)
return loss
class SmoothPenaltyLoss(k.losses.Loss):
def __init__(self, class_weights=np.ones(len(robot_label)), smooth=K.epsilon(), penalty=2.5,
name="smooth_penalty_loss"):
super(SmoothPenaltyLoss, self).__init__(name=name)
self.smooth = smooth
self.penalty = penalty
self.class_weights = tf.convert_to_tensor(class_weights, dtype=np.float32)
def call(self, y_true, y_pred):
# dice = dice_coef(y_true, y_pred, self.smooth)
n_classes = self.class_weights.shape[0]
dices = tf.convert_to_tensor([dice_coef(y_true[:, :, :, i], y_pred[:, :, :, i]) for i in range(n_classes)])
pgds = tf.convert_to_tensor([dices[i] / (1 + self.penalty * (1 - dices[i])) for i in range(n_classes)])
weighted_pgds = self.class_weights * pgds
return K.sum(self.class_weights) - K.sum(weighted_pgds)
class SmoothPenaltyLossSampleWise(k.losses.Loss):
def __init__(self, class_weights=[1., 1.], penalty=2.5, name="smooth_penalty_loss_sample_wise"):
super(SmoothPenaltyLossSampleWise, self).__init__(name=name)
self.penalty = penalty
self.class_weights = tf.convert_to_tensor(class_weights, dtype=np.float32)
def penalty_l(self, y_true, y_pred):
# dice = dice_coef(y_true, y_pred, self.smooth)
n_classes = self.class_weights.shape[0]
dices = tf.convert_to_tensor([dice_coef(y_true[:, :, i], y_pred[:, :, i]) for i in range(n_classes)])
pgds = tf.convert_to_tensor([dices[i] / (1 + self.penalty * (1 - dices[i])) for i in range(n_classes)])
weighted_pgds = self.class_weights * pgds
return K.sum(self.class_weights) - K.sum(weighted_pgds)
def call(self, y_true, y_pred):
per_sample_loss = tf.map_fn(lambda x: self.penalty_l(x[0], x[1]), (y_true, y_pred), dtype=tf.float32)
loss = tf.math.reduce_mean(per_sample_loss)
return loss
def iou_loss(y_true, y_pred):
y_true = tf.reshape(y_true, [-1])
y_pred = tf.reshape(y_pred, [-1])
intersection = tf.reduce_sum(y_true * y_pred)
union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) - intersection
score = (intersection + K.epsilon()) / (union + K.epsilon())
return 1 - score
@tf.autograph.experimental.do_not_convert
def dice_coef(y_true, y_pred, smooth=K.epsilon()):
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
dice = (2. * intersection + smooth) / (K.sum(y_true_f * y_true_f) + K.sum(y_pred_f * y_pred_f) + smooth)
return dice
def dice_loss_no_square(y_true, y_pred):
smooth = K.epsilon()
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
dice = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
return 1 - dice
def dice_loss_square(y_true, y_pred):
smooth = K.epsilon()
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
dice = (2. * intersection + smooth) / (K.sum(K.square(y_true_f)) + K.sum(K.square(y_pred_f)) + smooth)
return 1 - dice
def dice_coefficient_samplewise(y_true, y_pred):
smooth = K.epsilon()
intersection = K.sum(K.abs(y_true * y_pred), axis=(-1, -2, -3))
return (2. * intersection + smooth) / (
K.sum(K.square(y_true), axis=(-1, -2, -3)) + K.sum(K.square(y_pred), axis=(-1, -2, -3)) + smooth)
@tf.autograph.experimental.do_not_convert
def dice_coefficient_pixelwise(y_true, y_pred):
smooth = K.epsilon()
intersection = K.sum(K.abs(y_true * y_pred), axis=(-1, -4))
pixel_dice = (2. * intersection + smooth) / (
K.sum(K.square(y_true), axis=(-1, -4)) + K.sum(K.square(y_pred), axis=(-1, -4)) + smooth)
return 1 - K.mean(pixel_dice)
def dice_batch_plus_pixel(y_true, y_pred):
return dice_loss_square(y_true, y_pred) + dice_coefficient_pixelwise(y_true, y_pred) ** 5
def dice_sample_pixel(y_true, y_pred):
return dice_coefficient_samplewise_loss(y_true, y_pred) + dice_coefficient_pixelwise(y_true, y_pred) ** 5
@tf.autograph.experimental.do_not_convert
def dice_sample_pixel_ce(y_true, y_pred):
loss1 = dice_coefficient_samplewise_loss(y_true, y_pred)
# loss2 = dice_coefficient_pixelwise(y_true, y_pred)
# loss2 = loss2 ** 2
loss3 = tf.keras.metrics.binary_focal_crossentropy(y_true, y_pred)
loss3 = K.mean(loss3)
# return loss1 + loss3
return loss3
def focal(p_t, gamma=2):
return -(1 - p_t) ** gamma * K.log(p_t)
def log_thing(p_t, offset=0.02):
correction = K.log(1 + offset)
return -K.log(p_t + offset) + correction
@tf.autograph.experimental.do_not_convert
def dice_coefficient_samplewise_loss(y_true, y_pred):
dices = dice_coefficient_samplewise(y_true, y_pred)
# transformed_dices = log_thing(dices)
# transformed_dices = focal(dices, .5)
# transformed_dices = K.square(1-dices)
transformed_dices = 1 - dices
return tf.math.reduce_mean(transformed_dices)
def line_dice(y_true, y_pred):
return dice_coef(y_true[:, :, :, 0], y_pred[:, :, :, 0])
def self_dice(y_true, y_pred):
return dice_coef(y_true[:, :, :, 1], y_pred[:, :, :, 1])
def ball_dice(y_true, y_pred):
return dice_coef(y_true[:, :, :, 1], y_pred[:, :, :, 1])
def field_dice(y_true, y_pred):
return dice_coef(y_true[:, :, :, 3], y_pred[:, :, :, 3])
@tf.autograph.experimental.do_not_convert
def robot_dice(y_true, y_pred):
return dice_coefficient_samplewise(y_true[:, :, :, 0], y_pred[:, :, :, 0])
class WeightedTopKWorstLoss(k.losses.Loss):
def __init__(self, class_weights, k, name='weighted_top_k_worst_loss'):
super(WeightedTopKWorstLoss, self).__init__(name=name)
self.class_weights = class_weights
self.k = k
def top_k_worst_loss(self, y_true, y_pred):
per_pixel_loss = tf.nn.sigmoid_cross_entropy_with_logits(y_true, y_pred)
per_class_loss = tf.reduce_sum(per_pixel_loss, axis=[1, 2, 3]) # Sum over pixel dimensions
_, top_k_indices = tf.nn.top_k(per_class_loss, k=self.k) # Get indices of top-k worst pixels
top_k_indices = tf.expand_dims(top_k_indices, axis=-1)
top_k_true = tf.gather(y_true, top_k_indices, axis=1) # Gather top-k worst pixels for each class
top_k_pred = tf.gather(y_pred, top_k_indices, axis=1)
weighted_per_class_loss = tf.reduce_sum(top_k_true * tf.nn.sigmoid_cross_entropy_with_logits(top_k_true, top_k_pred), axis=[1, 2, 3])
weighted_loss = tf.reduce_mean(weighted_per_class_loss * self.class_weights)
return weighted_loss
def call(self, y_true, y_pred):
y_true = tf.cast(y_true, y_pred.dtype)
return self.top_k_worst_loss(y_true, y_pred)
def top_k_worst_loss(y_true, y_pred):
per_pixel_loss = K.binary_crossentropy(y_true, y_pred, from_logits=False)
per_class_loss = tf.reduce_sum(per_pixel_loss, axis=[1, 2, 3]) # Sum over pixel dimensions
_, top_k_indices = tf.nn.top_k(per_class_loss, k=10) # Get indices of top-k worst pixels
print(top_k_indices)
top_k_indices = tf.expand_dims(top_k_indices, axis=-1)
top_k_true = tf.gather(y_true, top_k_indices, axis=1) # Gather top-k worst pixels for each class
top_k_pred = tf.gather(y_pred, top_k_indices, axis=1)
weighted_per_class_loss = tf.reduce_sum(top_k_true * tf.nn.sigmoid_cross_entropy_with_logits(top_k_true, top_k_pred), axis=[1, 2, 3])
weighted_loss = tf.reduce_mean(weighted_per_class_loss * self.class_weights)