-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
310 lines (260 loc) · 12.6 KB
/
losses.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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
import keras.backend as K
import numpy as np
def sigmoid_np(val):
return 1/(1+np.exp(-val))
def iou_np(y_true, y_pred, eps=1e-6):
y_t_flat = y_true.flatten()
y_p_flat = y_pred.flatten()
intersection = np.sum(y_t_flat * y_p_flat)
union = np.sum(y_t_flat) + np.sum(y_p_flat) - intersection
return np.mean( (intersection + eps) / (union + eps))
def iou(y_true, y_pred, eps=1e-6):
y_t_flat = K.flatten(y_true)
y_p_flat = K.flatten(y_pred)
intersection = K.sum(y_t_flat * y_p_flat)
union = K.sum(y_t_flat) + K.sum(y_p_flat) - intersection
return K.mean( (intersection + eps) / (union + eps))
def iou_loss(y_true, y_pred):
return 1. - iou(y_true, y_pred)
class bce():
def __init__(self, use_loss_weights=False):
self.__name__ = 'bce'
self.use_loss_weights = use_loss_weights
def __call__(self, y_true, y_pred):
if not self.use_loss_weights:
return K.mean(K.binary_crossentropy(y_true, y_pred))
else:
y_loss_weights = K.expand_dims(y_true[...,1], axis=-1)
y_true = K.expand_dims(y_true[...,0], axis=-1)
return K.mean(y_loss_weights * K.binary_crossentropy(y_true, y_pred))
class bce_np():
def __init__(self, use_loss_weights=False, scalar_loss=True):
self.__name__ = 'bce'
self.use_loss_weights = use_loss_weights
self.scalar_loss=scalar_loss
def __call__(self, y_true, y_pred):
if self.use_loss_weights:
y_loss_weights = np.expand_dims(y_true[...,1], axis=-1)
y_true = np.expand_dims(y_true[...,0], axis=-1)
# taken from numpy_backend.py in Keras/backend
y_pred_clip = np.clip(y_pred, 1e-7, 1-1e-7)
y_pred_clip = np.log(y_pred_clip / (1 - y_pred_clip))
loss = y_true * -np.log(sigmoid_np(y_pred_clip)) + (1 - y_true) * -np.log(1 - sigmoid_np(y_pred_clip))
if self.use_loss_weights:
loss = loss * y_loss_weights
if self.scalar_loss:
return np.mean(loss)
else:
return np.mean(loss, axis=tuple(range(1,loss.ndim)))
class dice_coef():
def __init__(self, smooth=1, use_loss_weights=False, ignore_loss=False):
self.smooth = smooth
self.__name__ = 'sdice'
self.use_loss_weights = use_loss_weights
self.ignore_loss = ignore_loss
def __call__(self, y_true, y_pred):
if not self.use_loss_weights:
y_t_flat = K.flatten(y_true)
y_p_flat = K.flatten(y_pred)
intersection = K.sum(y_t_flat * y_p_flat)
union = K.sum(y_t_flat) + K.sum(y_p_flat)
return (2. * intersection + self.smooth) / (union + self.smooth)
elif not self.ignore_loss:
y_loss_weights = K.expand_dims(y_true[...,1], axis=-1)
y_true = K.expand_dims(y_true[...,0], axis=-1)
y_t_flat = K.flatten(y_true)
y_p_flat = K.flatten(y_pred)
y_lw_flat = K.flatten(y_loss_weights)
intersection = K.sum(y_t_flat * y_p_flat * y_lw_flat)
union = K.sum(y_t_flat * y_lw_flat) + K.sum(y_p_flat * y_lw_flat)
return (2. * intersection + self.smooth) / (union + self.smooth)
else:
y_loss_weights = K.expand_dims(y_true[...,1], axis=-1)
y_true = K.expand_dims(y_true[...,0], axis=-1)
y_t_flat = K.flatten(y_true)
y_p_flat = K.flatten(y_pred)
intersection = K.sum(y_t_flat * y_p_flat)
union = K.sum(y_t_flat) + K.sum(y_p_flat)
return (2. * intersection + self.smooth) / (union + self.smooth)
class dice_coef_np():
def __init__(self, smooth=1, use_loss_weights=False, ignore_loss=False, scalar_loss=True):
self.smooth = smooth
self.__name__ = 'sdice'
self.use_loss_weights = use_loss_weights
self.ignore_loss = ignore_loss
self.scalar_loss = scalar_loss
def __call__(self, y_true, y_pred):
if not self.use_loss_weights:
intersection = np.sum(y_true * y_pred)
union = np.sum(y_true) + np.sum(y_pred)
loss = (2. * intersection + self.smooth) / (union + self.smooth)
elif not self.ignore_loss:
y_loss_weights = np.expand_dims(y_true[...,1], axis=-1)
y_true = np.expand_dims(y_true[...,0], axis=-1)
intersection = np.sum(y_true * y_pred * y_loss_weights)
union = np.sum(y_true * y_loss_weights) + np.sum(y_pred * y_loss_weights)
loss = (2. * intersection + self.smooth) / (union + self.smooth)
else:
y_true = np.expand_dims(y_true[...,0], axis=-1)
intersection = np.sum(y_true * y_pred)
union = np.sum(y_true) + np.sum(y_pred)
loss = (2. * intersection + self.smooth) / (union + self.smooth)
if self.scalar_loss:
return np.mean(loss)
else:
return np.mean(loss, axis=tuple(range(1,loss.ndim)))
class dice_coef_loss():
def __init__(self, smooth=1, use_loss_weights=False, ignore_loss=True):
self.dice_coef=dice_coef(smooth=smooth, use_loss_weights=use_loss_weights, ignore_loss=True)
self.__name__ = 'sdice_l'
def __call__(self, y_true, y_pred):
return 1 - self.dice_coef(y_true, y_pred)
class dice_coef_loss_np():
def __init__(self, smooth=1, use_loss_weights=False, ignore_loss=True, scalar_loss=False):
self.dice_coef=dice_coef_np(smooth=smooth, use_loss_weights=use_loss_weights, ignore_loss=True)
self.__name__ = 'sdice_l'
def __call__(self, y_true, y_pred):
return 1 - self.dice_coef(y_true, y_pred)
import tensorflow as tf
import sys
class focal_loss():
def __init__(self, gamma=2., alpha=.25, use_loss_weights=False, ignore_loss=True):
self.gamma = gamma
self.alpha = alpha
self.__name__ = 'foc2_l'
self.use_loss_weights=use_loss_weights
self.ignore_loss=ignore_loss
def __call__(self, y_true, y_pred):
if self.use_loss_weights:
y_loss_weights = K.expand_dims(y_true[...,1], axis=-1)
y_true = K.expand_dims(y_true[...,0], axis=-1)
'''
pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
return -K.sum(self.alpha * K.pow(1. - pt_1, self.gamma) * K.log(pt_1))-K.sum((1-self.alpha) * K.pow( pt_0, self.gamma) * K.log(1. - pt_0))
'''
max_val = K.clip(-y_pred, min_value = 0, max_value = sys.float_info.max)
loss = y_pred - y_pred * y_true + max_val + K.log(K.exp(-max_val) + K.exp(-y_pred - max_val))
# This formula gives us the log sigmoid of 1-p if y is 0 and of p if y is 1
invprobs = K.log(K.sigmoid(-y_pred * (y_true * 2 - 1)))
loss = K.exp(invprobs * self.gamma) * loss
if not self.use_loss_weights:
return K.mean(loss)
elif not self.ignore_loss:
return K.mean(loss * y_loss_weights)
else:
return K.mean(loss)
class focal_loss_np():
def __init__(self, gamma=2., alpha=.25, use_loss_weights=False, ignore_loss=True, scalar_loss=True):
self.gamma = gamma
self.alpha = alpha
self.__name__ = 'foc2_l'
self.use_loss_weights=use_loss_weights
self.ignore_loss=ignore_loss
self.scalar_loss=scalar_loss
def __call__(self, y_true, y_pred):
if self.use_loss_weights:
y_loss_weights = np.expand_dims(y_true[...,1], axis=-1)
y_true = np.expand_dims(y_true[...,0], axis=-1)
'''
pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
return -K.sum(self.alpha * K.pow(1. - pt_1, self.gamma) * K.log(pt_1))-K.sum((1-self.alpha) * K.pow( pt_0, self.gamma) * K.log(1. - pt_0))
'''
max_val = np.clip(-y_pred, a_min = 0, a_max = sys.float_info.max)
loss = y_pred - y_pred * y_true + max_val + np.log(np.exp(-max_val) + np.exp(-y_pred - max_val))
# This formula gives us the log sigmoid of 1-p if y is 0 and of p if y is 1
invprobs = np.log(sigmoid_np(-y_pred * (y_true * 2 - 1)))
loss = np.exp(invprobs * self.gamma) * loss
if self.use_loss_weights and not self.ignore_loss:
loss = loss * y_loss_weights
if self.scalar_loss:
return np.mean(loss)
else:
return np.mean(loss, axis=tuple(range(1,loss.ndim)))
class mixed_loss():
def __init__(self, alpha=0.25, gamma=2., beta=1000., kappa=100., normalize=False, use_loss_weights=False):
self.alpha = alpha
self.gamma = gamma
self.beta = beta
self.kappa = kappa
self.loss_focal = focal_loss(gamma=gamma, alpha=alpha, use_loss_weights=use_loss_weights)
self.loss_dice = dice_coef_loss(use_loss_weights=use_loss_weights)
self.loss_tp = tp_loss()
self.__name__ = 'mixed_l'
def __call__(self, y_true, y_pred):
#return self.loss_focal(y_true, y_pred) + self.beta * self.loss_dice(y_true, y_pred) + self.kappa * self.loss_tp(y_true, y_pred)
return self.beta * self.loss_focal(y_true, y_pred) + self.kappa * self.loss_dice(y_true, y_pred)
class mixed_loss_np():
def __init__(self, alpha=0.25, gamma=2., beta=1000., kappa=100., normalize=False, use_loss_weights=False):
self.alpha = alpha
self.gamma = gamma
self.beta = beta
self.kappa = kappa
self.loss_focal = focal_loss_np(gamma=gamma, alpha=alpha, use_loss_weights=use_loss_weights, scalar_loss=False)
self.loss_dice = dice_coef_loss_np(use_loss_weights=use_loss_weights, scalar_loss=False)
self.loss_tp = tp_loss()
self.__name__ = 'mixed_l'
def __call__(self, y_true, y_pred):
#return self.loss_focal(y_true, y_pred) + self.beta * self.loss_dice(y_true, y_pred) + self.kappa * self.loss_tp(y_true, y_pred)
return self.beta * self.loss_focal(y_true, y_pred) + self.kappa * self.loss_dice(y_true, y_pred)
class mixed_loss2():
def __init__(self, alpha=0.25, gamma=2., beta=1000., kappa=100., normalize=False, use_loss_weights=False):
self.alpha = alpha
self.gamma = gamma
self.beta = beta
self.kappa = kappa
self.loss_focal = focal_loss(gamma=gamma, alpha=alpha, use_loss_weights=use_loss_weights)
self.loss_bce = bce(use_loss_weights=use_loss_weights)
self.loss_tp = tp_loss()
self.__name__ = 'mixed_l'
def __call__(self, y_true, y_pred):
return self.beta * self.loss_focal(y_true, y_pred) + self.kappa * self.loss_bce(y_true, y_pred)
class mixed_loss2_np():
def __init__(self, alpha=0.25, gamma=2., beta=1000., kappa=100., normalize=False, use_loss_weights=False, scalar_loss=True):
self.alpha = alpha
self.gamma = gamma
self.beta = beta
self.kappa = kappa
self.loss_focal = focal_loss_np(gamma=gamma, alpha=alpha, use_loss_weights=use_loss_weights, scalar_loss=scalar_loss)
self.loss_bce = bce_np(use_loss_weights=use_loss_weights, scalar_loss=scalar_loss)
self.__name__ = 'mixed_l'
def __call__(self, y_true, y_pred):
return self.beta * self.loss_focal(y_true, y_pred) + self.kappa * self.loss_bce(y_true, y_pred)
class true_positive_rate():
def __init__(self, eps=1e-6):
self.eps = eps
self.__name__ = 'tp'
def __call__(self, y_true, y_pred):
y_t_flat = K.flatten(y_true)
y_p_flat = K.flatten(y_pred)
return K.sum(y_t_flat * y_p_flat + self.eps)/K.sum(y_t_flat + self.eps)
class tp_loss():
def __init__(self, eps=1e-6):
self.tp = true_positive_rate(eps)
self.__name__ = 'tp_l'
def __call__(self, y_true, y_pred):
return 1 - self.tp(y_true, y_pred)
def dice_tp_loss(y_true, y_pred, alpha=0.999):
return alpha * dice_coef_loss(y_true, y_pred) + (1-alpha) * tp_loss(y_true, y_pred)
thresholds = [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95]
def f2(masks_true, masks_pred):
if np.sum(masks_true) == 0:
return float(np.sum(masks_pred) == 0)
ious = []
mp_idx_found = []
for mt in masks_true:
for mp_idx, mp in enumerate(masks_pred):
if mp_idx not in mp_idx_found:
cur_iou = iou_np(mt,mp)
if cur_iou > 0.5:
ious.append(cur_iou)
mp_idx_found.append(mp_idx)
break
f2_total = 0
for th in thresholds:
tp = sum([iou > th for iou in ious])
fn = len(masks_true) - tp
fp = len(masks_pred) - tp
f2_total += (5*tp)/(5*tp + 4*fn + fp)
return f2_total/len(thresholds)