-
Notifications
You must be signed in to change notification settings - Fork 5
/
core.pyx
214 lines (185 loc) · 6.09 KB
/
core.pyx
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
cimport cython
from config import Config
from cython cimport bool
import numpy as np
from cmath import e
from utils import show_status
# parameters
class LRError(Exception):
pass
class DataSpace(object):
def __init__(self):
self.B = 0.5 # rand b #TODO single?
self.C = 0.5 # rand 0-1 b_j
self.W = np.random.rand(Config.J) # rand 0-1 w_i
self.W_ = np.random.rand(Config.J, Config.K) # rand 0-1 w_jk
def tofile(self, path):
"""
put model parameters to file
"""
show_status(".. put parameters to file: %s" % path)
with open(path, 'w') as f:
tem = []
# input config K, J
tem.append(' '.join([str(i)
for i in (Config.K, Config.J)]))
# input config THETA SIGMA
tem.append(' '.join([str(i)
for i in (Config.THETA, Config.SIGMA)]))
# input parameters
# input a line of B, C
tem.append(' '.join([str(i)
for i in (self.B, self.C)]))
# input a line of Ws
tem.append(' '.join([str(i) for i in self.W]))
# input K lines of W_s
for j in range(Config.J):
tem.append(' '.join([str(i) for i in self.W_[j]]))
f.write('\n'.join(tem))
def fromfile(self, path):
show_status(".. load parameters from file : %s" % path)
def split_line_trans_type(line, _type):
return [_type(i) for i in line.split()]
with open(path) as f:
for no, line in enumerate(f.readlines()):
if no == 0:
Config.K, Config.J = split_line_trans_type(line, int)
elif no == 1:
Config.THETA, Config.SIGMA = split_line_trans_type(line, float)
elif no == 2:
self.B, self.C = split_line_trans_type(line, float)
elif no == 3:
Ws = split_line_trans_type(line, float)
for i in range(Config.J):
self.W[i] = Ws[i]
else:
j = no - 3
w_j = split_line_trans_type(line, float)
for k in range(Config.K):
self.W_[j][k] = w_j[k]
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.cdivision(True)
cdef class LambdaRank:
cdef:
object dataspace
object X1, X2
object S1, S2
object _diff_L
def __cinit__(self):
self.dataspace = DataSpace()
def init(self, X1, X2):
"""
init pair (X1, X2)
"""
self.X1, self.X2 = X1, X2
self.S1, self.S2 = self.s(self.X1), self.s(self.X2)
self._diff_L = self.diff_L(None, None, self.S1, self.S2)
cdef float diff_L_b(self):
"""
\partial L
-----------
\partial b
"""
res = \
self._diff_L * \
(
sum([self.dataspace.W[j] *
self.diff_f(j, self.X1) for j in range(Config.J)])
-
sum([self.dataspace.W[j] *
self.diff_f(j, self.X2) for j in range(Config.J)]))
return res
cdef float diff_L_w_(self, int j):
"""
\partial L
-----------
\partial w_j
"""
return self._diff_L *\
(self.f_(j, self.X1) - self.f_(j, self.X2))
cdef float diff_L_w__(self, int j, int k):
"""
\partial L
-----------
\partial W_jk
"""
return self._diff_L * self.dataspace.W[j] * (
self.diff_f(j, self.X1) * self.X1[k]
-
self.diff_f(j, self.X2) * self.X2[k]
)
# ---------------updater ----------
cdef void update_w_(self, int j):
"""
update w_k
"""
self.dataspace.W[j] = self.dataspace.W[j] - \
Config.THETA * self.diff_L_w_(j)
cdef void update_w__(self, int j, int k):
self.dataspace.W_[j][k] = self.dataspace.W_[j][k] - \
Config.THETA * self.diff_L_w__(j, k)
cdef void update_b(self):
self.dataspace.B = self.dataspace.B - \
Config.THETA * self.diff_L_b()
# ----------------API---------------
def study_line(self, X1, X2):
"""
study from a line of record from trainset
"""
self.init(X1, X2)
for j in range(Config.J):
self.update_w_(j)
self.update_b()
for j in range(Config.J):
for k in range(Config.K):
self.update_w__(j, k)
def predict(self, X):
return self.s(X)
# ----------------private detail functions -----
cdef float J(self, object X, int j):
"""
\sum_k{w_{jk}x_{(k)} + b_j}
"""
return sum([self.dataspace.W_[j][k] * X[k]
for k in range(Config.K)]) + self.dataspace.B
cdef float f(self, float x):
"""
>>> f(2)
0.8807970779778823
"""
return 1/(1+e**(-x))
cdef float f_(self, int j, object X):
"""
f_j(X)
"""
return self.f(self.J(X, j))
cdef float s(self, object X):
"""
\sum_j{w_j f_j(J(X)) + b}
"""
return sum([self.dataspace.W[j] * self.f_(j, X)
for j in range(Config.J)]) + self.dataspace.C
cdef float diff_L(self, object X1, object X2, object S1=None, object S2=None):
"""
cal L'
"""
S1 = self.s(X1) if not S1 else S1
S2 = self.s(X2) if not S2 else S2
return - Config.SIGMA / (
1 + e ** (float(Config.SIGMA * (S1 - S2)))
)
cdef float diff_f(self, int j, object X, float JX=0.0):
"""
cal
\partial{f(D)}
--------------
\partial{D}
"""
cdef:
float _D
_D = self.J(X, j) if JX==0.0 else JX
return e**(-_D) / (1+e**(-_D))**2
if __name__ == '__main__':
import doctest
doctest.testmod()