-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathCustomLosses.py
215 lines (158 loc) · 7.52 KB
/
CustomLosses.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
"""
Contains my implementation of custom losses / validation functions.
Works with keras2.0 and tf1.11.
Author : Manohar Kuse <mpkuse@connect.ust.hk>
Created : 6th Nov, 2018
"""
from keras import backend as K
from keras.engine.topology import Layer
import keras
import numpy as np
#import cv2
import code
def triplet_loss2_maker( nP, nN, epsilon=0.3 ):
# As per the NetVLAD paper's words
# def triplet_loss2( params ):
# y_true, y_pred = params
def triplet_loss2( y_true, y_pred ):
""" Closed negative sample - farthest positive sample """
assert( y_pred.shape[1] == 1+nP+nN )
# y_pred.shape = shape=(?, 5, 512)
q = y_pred[:,0:1,:] # shape=(?, 1, 512)
P = y_pred[:,1:1+nP,:] # shape=(?, nP, 512)
N = y_pred[:,1+nP:,:] # shape=(?, nN, 512)
q_dot_P = keras.layers.dot( [q,P], axes=-1 ) # shape=(?, 1, nP)
q_dot_N = keras.layers.dot( [q,N], axes=-1 ) # shape=(?, 1, nN)
# epsilon = 0.3 # Your epsilon here
d_nearest_positive_sample = K.max( q_dot_P, axis=-1, keepdims=True )
S = q_dot_N - d_nearest_positive_sample + epsilon #difference between best +ve and all negatives.
return K.sum( K.maximum( 0., S ), axis=-1 )
return triplet_loss2
def allpair_hinge_loss_maker( nP, nN, epsilon=0.3 ):
# def allpair_hinge_loss( params ):
# y_true, y_pred = params
def allpair_hinge_loss(y_true, y_pred):
""" All pair loss """
# nP = 3
# nN = 2
assert( y_pred.shape[1] == 1+nP+nN )
# y_pred.shape = shape=(?, 5, 512)
q = y_pred[:,0:1,:] # shape=(?, 1, 512)
P = y_pred[:,1:1+nP,:] # shape=(?, 2, 512)
N = y_pred[:,1+nP:,:] # shape=(?, 2, 512)
q_dot_P = keras.layers.dot( [q,P], axes=-1 ) # shape=(?, 1, 2)
q_dot_N = keras.layers.dot( [q,N], axes=-1 ) # shape=(?, 1, 2)
# epsilon = 0.3 # Your epsilon here
zeros = K.zeros((nP, nN), dtype='float32')
ones_m = K.ones((nP,1), dtype='float32')
ones_n = K.ones((nN,1), dtype='float32')
_1m__qdotN_T = ones_m[None,:] * q_dot_N # 1m ( \delta^q_N )^T
qdotP__1n_T = K.permute_dimensions( ones_n[None,:] * q_dot_P, [0,2,1] ) # ( \delta^q_P ) 1n^T
_1m__1n_T = epsilon * ones_m[None,:] * K.permute_dimensions( ones_n[None,:], [0,2,1] ) # 1m 1n^T
aux = _1m__qdotN_T - qdotP__1n_T + _1m__1n_T
return K.sum( K.maximum(zeros, aux) , axis=[-1,-2] )
return allpair_hinge_loss
def allpair_count_goodfit_maker( nP, nN, epsilon=0.3 ):
# def allpair_count_goodfit( params ):
# y_true, y_pred = params
def allpair_count_goodfit(y_true, y_pred):
# nP = 3
# nN = 2
assert( y_pred.shape[1] == 1+nP+nN )
# y_pred.shape = shape=(?, 5, 512)
q = y_pred[:,0:1,:] # shape=(?, 1, 512)
P = y_pred[:,1:1+nP,:] # shape=(?, 2, 512)
N = y_pred[:,1+nP:,:] # shape=(?, 2, 512)
q_dot_P = keras.layers.dot( [q,P], axes=-1 ) # shape=(?, 1, 2)
q_dot_N = keras.layers.dot( [q,N], axes=-1 ) # shape=(?, 1, 2)
# epsilon = 0.3 # Your epsilon here
zeros = K.zeros((nP, nN), dtype='float32')
ones_m = K.ones((nP,1), dtype='float32')
ones_n = K.ones((nN,1), dtype='float32')
_1m__qdotN_T = ones_m[None,:] * q_dot_N # 1m ( \delta^q_N )^T
qdotP__1n_T = K.permute_dimensions( ones_n[None,:] * q_dot_P, [0,2,1] ) # ( \delta^q_P ) 1n^T
_1m__1n_T = epsilon * ones_m[None,:] * K.permute_dimensions( ones_n[None,:], [0,2,1] ) # 1m 1n^T
aux = _1m__qdotN_T - qdotP__1n_T + _1m__1n_T
return K.sum( K.cast( K.less_equal( aux , 0), 'float32' ), axis=[-1,-2] ) #number of pairs which satisfy out of total nP*nN pairs
return allpair_count_goodfit
def positive_set_deviation_maker( nP, nN ):
# def positive_set_deviation( params ):
# y_true, y_pred = params
def positive_set_deviation(y_true, y_pred):
assert( y_pred.shape[1] == 1+nP+nN )
# y_pred.shape = shape=(?, 5, 512)
q = y_pred[:,0:1,:] # shape=(?, 1, 512)
P = y_pred[:,1:1+nP,:] # shape=(?, 2, 512)
N = y_pred[:,1+nP:,:] # shape=(?, 2, 512)
q_dot_P = keras.layers.dot( [q,P], axes=-1 ) # shape=(?, 1, nP)
q_dot_N = keras.layers.dot( [q,N], axes=-1 ) # shape=(?, 1, nN)
p_std = K.std( q_dot_P, axis=[-1,-2] )
return p_std
return positive_set_deviation
def allpair_hinge_loss_with_positive_set_deviation_maker( nP, nN, epsilon=0.3, opt_lambda=1.0 ):
# def allpair_hinge_loss_with_positive_set_deviation( params ):
# y_true, y_pred = params
def allpair_hinge_loss_with_positive_set_deviation(y_true, y_pred):
""" All pair loss with positive set deviation"""
# nP = 3
# nN = 2
assert( y_pred.shape[1] == 1+nP+nN )
# y_pred.shape = shape=(?, 5, 512)
q = y_pred[:,0:1,:] # shape=(?, 1, 512)
P = y_pred[:,1:1+nP,:] # shape=(?, 2, 512)
N = y_pred[:,1+nP:,:] # shape=(?, 2, 512)
q_dot_P = keras.layers.dot( [q,P], axes=-1 ) # shape=(?, 1, 2)
q_dot_N = keras.layers.dot( [q,N], axes=-1 ) # shape=(?, 1, 2)
# epsilon = 0.3 # Your epsilon here
zeros = K.zeros((nP, nN), dtype='float32')
ones_m = K.ones((nP,1), dtype='float32')
ones_n = K.ones((nN,1), dtype='float32')
_1m__qdotN_T = ones_m[None,:] * q_dot_N # 1m ( \delta^q_N )^T
qdotP__1n_T = K.permute_dimensions( ones_n[None,:] * q_dot_P, [0,2,1] ) # ( \delta^q_P ) 1n^T
_1m__1n_T = epsilon * ones_m[None,:] * K.permute_dimensions( ones_n[None,:], [0,2,1] ) # 1m 1n^T
aux = _1m__qdotN_T - qdotP__1n_T + _1m__1n_T
p_std = K.std( q_dot_P, axis=[-1,-2] ) #positive_set_deviation term
return K.sum( K.maximum(zeros, aux) , axis=[-1,-2] ) + opt_lambda * p_std
return allpair_hinge_loss_with_positive_set_deviation
# Verify loss function
if __name__ == '__main__':
np.random.seed(0)
nP = 3
nN = 2
y_true = keras.layers.Input( shape=(nP+nN+1,7) )
y_pred = keras.layers.Input( shape=(nP+nN+1,7) )
w = keras.layers.Lambda( allpair_hinge_loss_with_positive_set_deviation_maker( nP=nP, nN=nN, epsilon=0.3, opt_lambda=1.0) )( [y_true, y_pred] )
# w = keras.layers.Lambda( allpair_hinge_loss_maker( nP=nP, nN=nN, epsilon=0.1) )( [y_true, y_pred] )
# w_c = keras.layers.Lambda( allpair_count_goodfit_maker( nP=nP, nN=nN, epsilon=0.1) )( [y_true, y_pred] )
# w_c = keras.layers.Lambda( positive_set_deviation_maker( nP=nP, nN=nN, opt_lambda=.5) )( [y_true, y_pred] )
model = keras.models.Model( inputs=[y_true,y_pred], outputs=[w] )
model.summary()
keras.utils.plot_model( model, show_shapes=True )
a = np.zeros( (10,6,7) ) #. this doesn't appear in the loss function as my loss functions are weakly supervised. don't care, this is y_true
b_ = np.zeros( (10,6,7) )
b = b_[0,:,:]
b = np.round( np.random.random( (6,7) ), 2)
b = b / np.linalg.norm( b, axis=1, keepdims=True )
# b[0,0] = (3./5)
# b[0,6] = (4./5)
#
# b[1,0] = (3./5)
# b[1,1] = (4./5)
#
# b[2,0] = (3./5)
# b[2,3] = (4./5)
#
# b[2,0] = (3./5)
# b[2,3] = (4./5)
#
# b[2,0] = (3./5)
# b[2,3] = (4./5)
# b[1:3,:] = np.round( np.random.random( (2,7)), 2 )
# b[3:,:] = np.round( np.random.random( (2,7)), 2 )
b_[0,:,:] = b
b_[2,:,:] = b
out = model.predict( [a,b_] )
aux = np.array( [[-0.05192798, -0.00773406],
[ 0.1755529 , 0.21974683],
[ 0.06959844, 0.11379236]])
quit()