-
Notifications
You must be signed in to change notification settings - Fork 0
/
mlossTest.py
158 lines (149 loc) · 5.11 KB
/
mlossTest.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
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Tue Nov 29 09:39:36 2016
@author: XFZ
"""
import mxnet as mx
from CusSigmoid import *
from math import isnan
def mLoss(in_data,M,D,aux):
alpha = 1
data = in_data[0]
labels = in_data[1].asnumpy()
wrapSize = M * D
batchSize = data.shape[0]
aux_f = aux[0]
aux_h = aux[1]
aux_diff = aux[2]
auxCentroid = aux[3]
auxSigma = aux[4]
wrapSize = M * D
batchSize = data.shape[0]
xpu = data.context
loss_out = mx.nd.zeros((batchSize,),ctx=xpu)
for ii in range(batchSize / wrapSize):
wrap=data[ii*wrapSize:(ii+1)*wrapSize]
wraplables = labels[ii*wrapSize:(ii+1)*wrapSize]
#calculate cluster centers
mu = []
for j in range(wrapSize/D):
c = mx.nd.sum(wrap[j*D:(j+1)*D],axis=0) / D
mu.append(c)
auxCentroid[ii*M+j][:]=c
#calculate diff
diff= []
for n in range(wrapSize):
diff.append([])
for m in range(M):
d = mx.nd.sum(mx.nd.square(wrap[n] - mu[m]))
#d = mx.nd.sqrt(d)
bug_d = d.asnumpy()
if (isnan(bug_d)):
import pdb;pdb.set_trace()
diff[n].append(d)
aux_diff[ii*wrapSize+n][m:m+1]=d
#calculate sigma
s=mx.nd.zeros(1,ctx = xpu)
k=0
m=0
for j in range(wrapSize):
s += diff[j][m]
k = k+1
if k>=D :
k=0
m+1
sigma = s.asnumpy()/(wrapSize - 1)
sigma = float(2*sigma[0])
auxSigma[ii:ii+1] = s
#calculate loss for wrap
loss=mx.nd.zeros(1,ctx=xpu)
frac = mx.nd.zeros(1,ctx=xpu)
for j in range(wrapSize):
frac[:] = 0
for i in range(M):
if wraplables[j] !=wraplables[i*D]:
frac += mx.nd.exp(- diff[j][i]/sigma)
aux_h[ii*wrapSize+j:ii*wrapSize+j+1] = frac
f=diff[j][int(j/D)]/sigma+alpha
loss += f + mx.nd.log(frac)
loss = loss / wrapSize
mx.nd.broadcast_to( loss,\
out=loss_out[ii*wrapSize:(ii+1)*wrapSize],\
shape = (wrapSize))
return loss_out
def get_mGrad( out_data,in_data,M,D, aux):
aux_f = aux[0]
aux_h = aux[1]
aux_d = aux[2]
auxCentroid = aux[3]
auxSigma = aux[4]
data = in_data[0]
xpu = data.context
labels = in_data[1].asnumpy()
M = int(M)
D = int (D)
wrapSize = M * D
batchSize = data.shape[0]
featureSize = data.shape[1]
grad = mx.nd.zeros((batchSize,featureSize),ctx = xpu)
Sigma= auxSigma.asnumpy()
part = mx.nd.zeros((featureSize,),ctx=xpu)
for ii in range(batchSize / wrapSize):
wrap=data[ii*wrapSize:(ii+1)*wrapSize]
wraplables = labels[ii*wrapSize:(ii+1)*wrapSize]
sigma = Sigma[ii]
for j in range(wrapSize):
part[:] = 0
cnt = 0
for i in range(M):
if wraplables[j] !=wraplables[i*D]:
score = mx.nd.exp(- aux_d[ii*wrapSize+j][i:i+1]/(2*sigma))
part += mx.nd.broadcast_mul(auxCentroid[ii*M+i],score)
cnt +=1
gh = mx.nd.broadcast_div(part,aux_h[ii*wrapSize+j:ii*wrapSize+j+1])
gf = wrap[j]-auxCentroid[ii*M+int(j/D)]
g = gf - wrap[j]*cnt +gh
g = g/(M * D * sigma)
#grad[ii*wrapSize+j] = g*y[ii*wrapSize+j]
grad[ii*wrapSize+j] = g
return grad
def Sbackward(out_grad, out_data,l):
y = out_data[0]
z = out_grad[0]
grad = y*(1 - y)
grad = grad * 2 *l
return z*grad
ctx = mx.gpu()
data=mx.random.normal(0,5,shape=(48,100),ctx=ctx)
ans =mx.random.normal(0,5,shape=(48,100),ctx=ctx)
nplabel = [0,0,0,0,0,0,0,0,1,1,1,1]
label = mx.nd.array([0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,0,0,1,1,0,0,3,3,\
0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,0,0,1,1,0,0,3,3],ctx=ctx)
f=mx.nd.zeros((48,),ctx = ctx)
d=mx.nd.zeros((48,12),ctx = ctx)
h=mx.nd.zeros((48,),ctx=ctx)
c=mx.nd.zeros((24,100),ctx=ctx)
s=mx.nd.zeros((2,),ctx=ctx)
aux = [f,h,d,c,s]
loss = mLoss([data,label],12,2,aux)
g = get_mGrad([loss],[data,label],12,2,aux)
sgrad = Sbackward([data],[data],3)
X = mx.symbol.Variable(name='data')
sigmoid = mx.symbol.Custom(data=X, name='sigmoid', op_type='Sigmoid',l=1)
Y = 3*sigmoid
Z = 4*sigmoid
F = Y + 0.5*Z
x = mx.random.normal(0,5,shape=(4,100))
y = mx.random.normal(0,5,shape=(4,100))
z = mx.random.normal(0,5,shape=(4,100))
f = mx.nd.ones((4,100))*2
gx = mx.nd.zeros((4,100))
gy = mx.nd.zeros((4,100))
gf = mx.nd.zeros((4,100))
gz = mx.nd.zeros((4,100))
ex = F.bind(ctx= ctx,args = {'data':f,'sigmoid':y},\
args_grad = {'data':gx,'sigmoid':gy} )
ex.forward()
out = ex.outputs[0]
ex.backward(out)