-
Notifications
You must be signed in to change notification settings - Fork 0
/
logreg_transfer.py
executable file
·160 lines (132 loc) · 4.82 KB
/
logreg_transfer.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
'''
Utilitarian functions
Dropout Classification Pipeline
'''
import csv
import sys
import numpy as np
from scipy.optimize import minimize
from sklearn import metrics
from sklearn.metrics import accuracy_score
import mpmath as mp
def compute_weight(X,Y,u,s,coef_transfer,e):
print "Computing weights"
b=generate_b(np.shape(X)[1]+1,e)
func=lambda W: logReg_ObjectiveFunction(X,Y,W,u,s,coef_transfer,b)
w=np.zeros((np.shape(X)[1]+1,1))
sol=minimize(func,w)
return sol.x
def generate_b(d,e):
res=np.random.randn(d)
norm=np.random.gamma(d,2/float(e))
s=sum(res)
res=res*(norm/s)
return np.array([res]).T
def logReg_ObjectiveFunction(X,Y,W,u,s,coef_transfer,b):
w=np.array([W[1:]]).T
w_0=W[0]
A=np.dot(X,w)+float(w_0)
B=-np.multiply(Y,A)
with mp.workdps(30):
NLL=sum(log_one_plus_exp(B))+coef_transfer*0.5*GPriorWeightRegTerm(W,u,s)#+(1/float(np.shape(Y)[0]))*np.dot(b.T,np.array([W]).T)[0,0]
return NLL
def log_one_plus_exp(B):
#overflow protection
#very slow, in future figure out a way to do this better with np matrices
n,m = np.shape(B)
output = np.zeros((n,m))
low_val = np.log(1)
for i,x in enumerate(B):
if x < -37:
output[i] = low_val
elif x > 37:
output[i] = x
else:
output[i] = np.log(1+np.exp(x))
return output
def GPriorWeightRegTerm(w,u,s):
result=0
for i in range(0,np.shape(u)[0]):
result+=((w[i]-u[i])/s[i])**2
return result
def estimatePrior(W):#W contains the list of w's computed for several task (row = weights of one task)
K=np.shape(W)[0] #number of tasks
u=(1/float(K))*np.sum(W,axis=0)
W_norm=W-u
s=np.sqrt((1/float(K-1))*np.sum(np.multiply(W_norm,W_norm),axis=0))
return u,s #### WRONG ANSWER ======> TO MODIFY
def separateAndComputeWeight(X,Y,u,s,n_tasks,coef_transfer):
n_sample=np.shape(X)[0]
n_sampleTask=int(n_sample/n_tasks)
W=np.zeros((n_tasks,1+np.shape(X)[1]))
for i in range(0,n_tasks):
X_task=X[i*n_sampleTask:(i+1)*n_sampleTask,:]
Y_task=Y[i*n_sampleTask:(i+1)*n_sampleTask,:]
W[i,:]=compute_weight(X_task,Y_task,u,s,coef_transfer)
return W
def computeWeight_fromPreviousTask(X_taskA,Y_taskA,X_taskB,Y_taskB,s_prior_fact,n_tasks,coef_transfer): # Compute weights for X_tasksB using assumed similarity with taskA
n_feat=np.shape(X_taskA)[1]
u=np.zeros((n_feat+1,1))
s=s_prior_fact*np.ones((n_feat+1,1))
W=separateAndComputeWeight(X_taskA,Y_taskA,u,s,n_tasks,coef_transfer)
u,s=estimatePrior(W)
sol=compute_weight(X_taskB,Y_taskB,u,s,coef_transfer)
return sol
def computeAUC(w,X_test,Y_test):
w_1=w[1:]
w_0=w[0]
pred=sigmoid(np.dot(X_test,w_1)+w_0)
fpr, tpr, thresholds = metrics.roc_curve(Y_test, pred, pos_label=1)
return metrics.auc(fpr, tpr)
def computeBestAccuracy(w,X_test,Y_test):
w_1=w[1:]
w_0=w[0]
pred=sigmoid(np.dot(X_test,w_1)+w_0)
fpr, tpr, thresholds = metrics.roc_curve(Y_test, pred, pos_label=1)
P=len(Y_test[Y_test==1])
N=np.shape(Y_test)[0]-P
print "P=",P,"N=",N
def acc(P,N,TPR,FPR):
return (TPR*P+N*(1-FPR))/float(P+N)
accuracies=[acc(P,N,tpr[i],fpr[i]) for i in range(len(fpr))]
# print "thresholds[np.argmax(accuracies)]",thresholds[np.argmax(accuracies)]
return max(accuracies)
def compute_Apriori_Accuracy(Y_train,Y_test):
print "np.shape(Y_test)",np.shape(Y_test)
if sum(Y_train)>0.5*np.shape(Y_train)[0]:
res=len(Y_test[Y_test==1])/float(np.shape(Y_test)[0])
else:
res=1-len(Y_test[Y_test==1])/float(np.shape(Y_test)[0])
return res
def compute_reverseAUC(w,X_test,Y_test):
w_1=w[1:]
w_0=w[0]
pred=1-sigmoid(np.dot(X_test,w_1)+w_0)
fpr, tpr, thresholds = metrics.roc_curve(Y_test, pred, pos_label=1)
return metrics.auc(fpr, tpr)
def computeAccuracy(w,X_test,Y_test):
w_1=w[1:]
w_0=w[0]
pred= np.dot(X_test,w_1)+w_0#-np.ones((1,np.shape(X_test)[0])).T#np.dot(X_test,w_1)+w_0#
print "prediction : ",pred
print "Number of positive prediction (dp=0)",sum([1 for x in pred if x >0])
result=[1 for x in np.multiply(pred,Y_test) if x>0]
acc=sum(result)/float(np.shape(X_test)[0])
return acc
def sigmoid(x):
return 1/(1+np.exp(-x))
def computeWeights_multiTask(X_A,X_B,Y_A,Y_B,l_particular,l_common):
print "Computing weights for multi task"
func=lambda W: Logreg_multitasks_objFunc(X_A,X_B,Y_A,Y_B,W,l_particular,l_common)
w=np.zeros((3*(np.shape(X_A)[1]+1),1))
sol=minimize(func,w)
return sol.x
# X_taskA=np.array([[1,2,2],[0,1,3],[3,3,0],[0,1,0]])
# Y_taskA=np.array([[1],[-1],[1],[-1]])
# X_taskB=np.array([[1,2,5],[0,1,3]])
# Y_taskB=np.array([[1],[-1]])
# W=np.zeros((np.shape(X_taskB)[1]+1,1))
# W=np.array([[0],[4],[0],[-1]])
# # u=np.zeros((np.shape(X_taskB)[1],1))
# # s=10*np.ones((np.shape(X_taskB)[1],1))
# print computeAccuracy(W,X_taskB,Y_taskB)