-
Notifications
You must be signed in to change notification settings - Fork 0
/
robsvm.py
332 lines (268 loc) · 11.4 KB
/
robsvm.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
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
from cvxopt import spmatrix, matrix, sparse, normal, mul, div, solvers, lapack, blas, base, misc, sqrt
def robsvm(X, d, gamma, P, e):
"""
Solves the following robust SVM training problem:
minimize (1/2) w'*w + gamma*sum(v)
subject to diag(d)*(X*w + b*1) >= 1 - v + E*u
|| S_j*w ||_2 <= u_j, j = 1...t
v >= 0
The variables are w, b, v, and u. The matrix E is a selector
matrix with zeros and one '1' per row. E_ij = 1 means that the
i'th training vector is associated with the j'th uncertainty
ellipsoid.
A custom KKT solver that exploits low-rank structure is used, and
a positive definite system of equations of order n is
formed and solved at each iteration.
ARGUMENTS
X m-by-n matrix with training vectors as rows
d m-vector with training labels (-1,+1)
P list of t symmetric matrices of order n
e m-vector where e[i] is the index of the uncertainty
ellipsoid associated with the i'th training vector
RETURNS
w n-vector
b scalar
u t-vector
v m-vector
iters number of interior-point iterations
"""
m,n = X.size
#m = X.shape[0]
#n = X.shape[1]
assert type(P) is list, "P must be a list of t symmtric positive definite matrices of order n."
k = len(P)
if k > 0:
assert e.size == (m,1), "e must be an m-vector."
assert max(e) < k and min(e) >= 0, "e[i] must be in {0,1,...,k-1}."
E = spmatrix(1.,e,range(m),(k,m)).T
d = matrix(d,tc='d')
q = matrix(0.0, (n+k+1+m,1))
q[n+k+1:] = gamma
h = matrix(0.0,(2*m+k*(n+1),1))
h[:m] = -1.0
# linear operators Q and G
def Q(x, y, alpha = 1.0, beta = 0.0, trans = 'N'):
y[:n] = alpha * x[:n] + beta * y[:n]
def G(x, y, alpha = 1.0, beta = 0.0, trans = 'N'):
"""
Implements the linear operator
[ -DX E -d -I ]
[ 0 0 0 -I ]
[ 0 -e_1' 0 0 ]
G = [ -P_1' 0 0 0 ]
[ . . . . ]
[ 0 -e_k' 0 0 ]
[ -P_k' 0 0 0 ]
and its adjoint G'.
"""
if trans == 'N':
tmp = +y[:m]
# y[:m] = alpha*(-DXw + Et - d*b - v) + beta*y[:m]
base.gemv(E, x[n:n+k], tmp, alpha = alpha, beta = beta)
blas.axpy(x[n+k+1:], tmp, alpha = -alpha)
blas.axpy(d, tmp, alpha = -alpha*x[n+k])
y[:m] = tmp
base.gemv(X, x[:n], tmp, alpha = alpha, beta = 0.0)
tmp = mul(d,tmp)
y[:m] -= tmp
# y[m:2*m] = -v
y[m:2*m] = -alpha * x[n+k+1:] + beta * y[m:2*m]
# SOC 1,...,k
for i in range(k):
l = 2*m+i*(n+1)
y[l] = -alpha * x[n+i] + beta * y[l]
y[l+1:l+1+n] = -alpha * P[i] * x[:n] + beta * y[l+1:l+1+n];
else:
tmp1 = mul(d,x[:m])
tmp2 = y[:n]
blas.gemv(X, tmp1, tmp2, trans = 'T', alpha = -alpha, beta = beta)
for i in range(k):
l = 2*m+1+i*(n+1)
blas.gemv(P[i], x[l:l+n], tmp2, trans = 'T', alpha = -alpha, beta = 1.0)
y[:n] = tmp2
tmp2 = y[n:n+k]
base.gemv(E, x[:m], tmp2, trans = 'T', alpha = alpha, beta = beta)
blas.axpy(x[2*m:2*m+k*(1+n):n+1], tmp2, alpha = -alpha)
y[n:n+k] = tmp2
y[n+k] = -alpha * blas.dot(d,x[:m]) + beta * y[n+k]
y[n+k+1:] = -alpha * (x[:m] + x[m:2*m]) + beta * y[n+k+1:]
# precompute products Pi'*Pi
Pt = []
for p in P:
y = matrix(0.0, (n,n))
blas.syrk(p, y, trans = 'T')
Pt.append(y)
# scaled hyperbolic Householder transformations
def qscal(u, beta, v, inv = False):
"""
Transforms the vector u as
u := beta * (2*v*v' - J) * u
if 'inv' is False and as
u := (1/beta) * (2*J*v*v'*J - J) * u
if 'inv' is True.
"""
if not inv:
tmp = blas.dot(u,v)
u[0] *= -1
u += 2 * v * tmp
u *= beta
else:
u[0] *= -1.0
tmp = blas.dot(v,u)
u[0] -= 2*v[0] * tmp
u[1:] += 2*v[1:] * tmp
u /= beta
# custom KKT solver
def F(W):
"""
Custom solver for the system
[ It 0 0 Xt' 0 At1' ... Atk' ][ dwt ] [ rwt ]
[ 0 0 0 -d' 0 0 ... 0 ][ db ] [ rb ]
[ 0 0 0 -I -I 0 ... 0 ][ dv ] [ rv ]
[ Xt -d -I -Wl1^-2 ][ dzl1 ] [ rl1 ]
[ 0 0 -I -Wl2^-2 ][ dzl2 ] = [ rl2 ]
[ At1 0 0 -W1^-2 ][ dz1 ] [ r1 ]
[ | | | . ][ | ] [ | ]
[ Atk 0 0 -Wk^-2 ][ dzk ] [ rk ]
where
It = [ I 0 ] Xt = [ -D*X E ] Ati = [ 0 -e_i' ]
[ 0 0 ] [ -Pi 0 ]
dwt = [ dw ] rwt = [ rw ]
[ dt ] [ rt ].
"""
# scalings and 'intermediate' vectors
# db = inv(Wl1)^2 + inv(Wl2)^2
db = W['di'][:m]**2 + W['di'][m:2*m]**2
dbi = div(1.0,db)
# dt = I - inv(Wl1)*Dbi*inv(Wl1)
dt = 1.0 - mul(W['di'][:m]**2,dbi)
dtsqrt = sqrt(dt)
# lam = Dt*inv(Wl1)*d
lam = mul(dt,mul(W['di'][:m],d))
# lt = E'*inv(Wl1)*lam
lt = matrix(0.0,(k,1))
base.gemv(E, mul(W['di'][:m],lam), lt, trans = 'T')
# Xs = sqrt(Dt)*inv(Wl1)*X
tmp = mul(dtsqrt,W['di'][:m])
Xs = spmatrix(tmp,range(m),range(m))*X
# Es = D*sqrt(Dt)*inv(Wl1)*E
Es = spmatrix(mul(d,tmp),range(m),range(m))*E
# form Ab = I + sum((1/bi)^2*(Pi'*Pi + 4*(v'*v + 1)*Pi'*y*y'*Pi)) + Xs'*Xs
# and Bb = -sum((1/bi)^2*(4*ui*v'*v*Pi'*y*ei')) - Xs'*Es
# and D2 = Es'*Es + sum((1/bi)^2*(1+4*ui^2*(v'*v - 1))
Ab = matrix(0.0,(n,n))
Ab[::n+1] = 1.0
base.syrk(Xs,Ab,trans = 'T', beta = 1.0)
Bb = matrix(0.0,(n,k))
Bb = -Xs.T*Es # inefficient!?
D2 = spmatrix(0.0,range(k),range(k))
base.syrk(Es,D2,trans = 'T', partial = True)
d2 = +D2.V
del D2
py = matrix(0.0,(n,1))
for i in range(k):
binvsq = (1.0/W['beta'][i])**2
Ab += binvsq*Pt[i]
dvv = blas.dot(W['v'][i],W['v'][i])
blas.gemv(P[i], W['v'][i][1:], py, trans = 'T', alpha = 1.0, beta = 0.0)
blas.syrk(py, Ab, alpha = 4*binvsq*(dvv+1), beta = 1.0)
Bb[:,i] -= 4*binvsq*W['v'][i][0]*dvv*py
d2[i] += binvsq*(1+4*(W['v'][i][0]**2)*(dvv-1))
d2i = div(1.0,d2)
d2isqrt = sqrt(d2i)
# compute a = alpha - lam'*inv(Wl1)*E*inv(D2)*E'*inv(Wl1)*lam
alpha = blas.dot(lam,mul(W['di'][:m],d))
tmp = matrix(0.0,(k,1))
base.gemv(E,mul(W['di'][:m],lam), tmp, trans = 'T')
tmp = mul(tmp, d2isqrt) #tmp = inv(D2)^(1/2)*E'*inv(Wl1)*lam
a = alpha - blas.dot(tmp,tmp)
# compute M12 = X'*D*inv(Wl1)*lam + Bb*inv(D2)*E'*inv(Wl1)*lam
tmp = mul(tmp, d2isqrt)
M12 = matrix(0.0,(n,1))
blas.gemv(Bb,tmp,M12, alpha = 1.0)
tmp = mul(d,mul(W['di'][:m],lam))
blas.gemv(X,tmp,M12, trans = 'T', alpha = 1.0, beta = 1.0)
# form and factor M
sBb = Bb * spmatrix(d2isqrt,range(k), range(k))
base.syrk(sBb, Ab, alpha = -1.0, beta = 1.0)
M = matrix([[Ab, M12.T],[M12, a]])
lapack.potrf(M)
def f(x,y,z):
# residuals
rwt = x[:n+k]
rb = x[n+k]
rv = x[n+k+1:n+k+1+m]
iw_rl1 = mul(W['di'][:m],z[:m])
iw_rl2 = mul(W['di'][m:2*m],z[m:2*m])
ri = [z[2*m+i*(n+1):2*m+(i+1)*(n+1)] for i in range(k)]
# compute 'derived' residuals
# rbwt = rwt + sum(Ai'*inv(Wi)^2*ri) + [-X'*D; E']*inv(Wl1)^2*rl1
rbwt = +rwt
for i in range(k):
tmp = +ri[i]
qscal(tmp,W['beta'][i],W['v'][i],inv=True)
qscal(tmp,W['beta'][i],W['v'][i],inv=True)
rbwt[n+i] -= tmp[0]
blas.gemv(P[i], tmp[1:], rbwt, trans = 'T', alpha = -1.0, beta = 1.0)
tmp = mul(W['di'][:m],iw_rl1)
tmp2 = matrix(0.0,(k,1))
base.gemv(E,tmp,tmp2,trans='T')
rbwt[n:] += tmp2
tmp = mul(d,tmp) # tmp = D*inv(Wl1)^2*rl1
blas.gemv(X,tmp,rbwt,trans='T', alpha = -1.0, beta = 1.0)
# rbb = rb - d'*inv(Wl1)^2*rl1
rbb = rb - sum(tmp)
# rbv = rv - inv(Wl2)*rl2 - inv(Wl1)^2*rl1
rbv = rv - mul(W['di'][m:2*m],iw_rl2) - mul(W['di'][:m],iw_rl1)
# [rtw;rtt] = rbwt + [-X'*D; E']*inv(Wl1)^2*inv(Db)*rbv
tmp = mul(W['di'][:m]**2, mul(dbi,rbv))
rtt = +rbwt[n:]
base.gemv(E, tmp, rtt, trans = 'T', alpha = 1.0, beta = 1.0)
rtw = +rbwt[:n]
tmp = mul(d,tmp)
blas.gemv(X, tmp, rtw, trans = 'T', alpha = -1.0, beta = 1.0)
# rtb = rbb - d'*inv(Wl1)^2*inv(Db)*rbv
rtb = rbb - sum(tmp)
# solve M*[dw;db] = [rtw - Bb*inv(D2)*rtt; rtb + lt'*inv(D2)*rtt]
tmp = mul(d2i,rtt)
tmp2 = matrix(0.0,(n,1))
blas.gemv(Bb,tmp,tmp2)
dwdb = matrix([rtw - tmp2,rtb + blas.dot(mul(d2i,lt),rtt)])
lapack.potrs(M,dwdb)
# compute dt = inv(D2)*(rtt - Bb'*dw + lt*db)
tmp2 = matrix(0.0,(k,1))
blas.gemv(Bb, dwdb[:n], tmp2, trans='T')
dt = mul(d2i, rtt - tmp2 + lt*dwdb[-1])
# compute dv = inv(Db)*(rbv + inv(Wl1)^2*(E*dt - D*X*dw - d*db))
dv = matrix(0.0,(m,1))
blas.gemv(X,dwdb[:n],dv,alpha = -1.0)
dv = mul(d,dv) - d*dwdb[-1]
base.gemv(E, dt, dv, beta = 1.0)
tmp = +dv # tmp = E*dt - D*X*dw - d*db
dv = mul(dbi, rbv + mul(W['di'][:m]**2,dv))
# compute wdz1 = inv(Wl1)*(E*dt - D*X*dw - d*db - dv - rl1)
wdz1 = mul(W['di'][:m], tmp - dv) - iw_rl1
# compute wdz2 = - inv(Wl2)*(dv + rl2)
wdz2 = - mul(W['di'][m:2*m],dv) - iw_rl2
# compute wdzi = inv(Wi)*([-ei'*dt; -Pi*dw] - ri)
wdzi = []
tmp = matrix(0.0,(n,1))
for i in range(k):
blas.gemv(P[i],dwdb[:n],tmp, alpha = -1.0, beta = 0.0)
tmp1 = matrix([-dt[i],tmp])
blas.axpy(ri[i],tmp1,alpha = -1.0)
qscal(tmp1,W['beta'][i],W['v'][i],inv=True)
wdzi.append(tmp1)
# solution
x[:n] = dwdb[:n]
x[n:n+k] = dt
x[n+k] = dwdb[-1]
x[n+k+1:] = dv
z[:m] = wdz1
z[m:2*m] = wdz2
for i in range(k):
z[2*m+i*(n+1):2*m+(i+1)*(n+1)] = wdzi[i]
return f
# solve cone QP and return solution
sol = solvers.coneqp(Q, q, G, h, dims = {'l':2*m,'q':[n+1 for i in range(k)],'s':[]}, kktsolver = F)
return sol['x'][:n], sol['x'][n+k], sol['x'][n:n+k], sol['x'][n+k+1:], sol['iterations']