def train_dc(self, max_iter=50):
""" Solve the optimization problem with a
sequential convex programming/DC-programming
approach:
Iteratively, find the most likely configuration of
the latent variables and then, optimize for the
model parameter using fixed latent states.
"""
N = self.sobj.get_num_samples()
DIMS = self.sobj.get_num_dims()
# intermediate solutions
# latent variables
latent = [0.0]*N
sol = normal(DIMS,1)
psi = matrix(0.0, (DIMS,N)) # (dim x exm)
old_psi = matrix(0.0, (DIMS,N)) # (dim x exm)
threshold = matrix(0.0)
obj = -1
iter = 0
# terminate if objective function value doesn't change much
while iter<max_iter and (iter<2 or sum(sum(abs(np.array(psi-old_psi))))>=0.001):
print('Starting iteration {0}.'.format(iter))
print(sum(sum(abs(np.array(psi-old_psi)))))
iter += 1
old_psi = matrix(psi)
# 1. linearize
# for the current solution compute the
# most likely latent variable configuration
mean = matrix(0.0, (DIMS, 1))
for i in range(N):
(foo, latent[i], psi[:,i]) = self.sobj.argmax(sol, i, add_prior=True)
mean += psi[:,i]
mpsi = matrix(psi)
mean /= float(N)
#for i in range(N):
#mphi[:,i] -= mean
# 2. solve the intermediate convex optimization problem
A = mpsi*mpsi.trans()
print A.size
W = matrix(0.0, (DIMS,DIMS))
syev(A,W,jobz='V')
print W
print A
print A*A.trans()
#sol = (W[3:,0].trans() * A[:,3:].trans()).trans()
#sol = (W[3:,0].trans() * A[:,3:].trans()).trans()
sol = A[:,DIMS-1]
print sol
print(sum(sum(abs(np.array(psi-old_psi)))))
self.sol = sol
self.latent = latent
return (sol, latent, threshold)
def using_numerical_derivatives(params=None, z=None):
out = _using_numerical_derivatives(params, z)
print "OUT",
for o in out:
print o
if len(out)==3:
from cvxopt.lapack import syev
W = cvx.matrix(0.0, (6,1), 'd')
syev(out[-1], W)
print "EIGEN", W
return out
def using_numerical_derivatives(params=None, z=None):
out = _using_numerical_derivatives(params, z)
print "OUT",
for o in out:
print o
if len(out) == 3:
from cvxopt.lapack import syev
W = cvx.matrix(0.0, (6, 1), 'd')
syev(out[-1], W)
print "EIGEN", W
return out
def fit(self, max_iter=50):
""" Solve the optimization problem with a
sequential convex programming/DC-programming
approach:
Iteratively, find the most likely configuration of
the latent variables and then, optimize for the
model parameter using fixed latent states.
"""
N = self.sobj.get_num_samples()
DIMS = self.sobj.get_num_dims()
# intermediate solutions
# latent variables
latent = [0.0] * N
sol = np.random.randn(DIMS)
psi = np.zeros((DIMS, N)) # (dim x exm)
old_psi = np.zeros((DIMS, N)) # (dim x exm)
threshold = 0.
obj = -1.
iter = 0
# terminate if objective function value doesn't change much
while iter < max_iter and (
iter < 2 or np.sum(abs(np.array(psi - old_psi))) >= 0.001):
print('Starting iteration {0}.'.format(iter))
print(np.sum(abs(np.array(psi - old_psi))))
iter += 1
old_psi = psi.copy()
# 1. linearize
# for the current solution compute the
# most likely latent variable configuration
mean = np.zeros(DIMS)
for i in range(N):
_, latent[i], psi[:, i] = self.sobj.argmax(sol, i)
mean += psi[:, i]
mean /= float(N)
mpsi = psi - np.repeat(mean.reshape((DIMS, 1)), N, axis=1)
# 2. solve the intermediate convex optimization problem
A = mpsi.dot(mpsi.T)
print A.shape
W = np.zeros((DIMS, DIMS))
syev(matrix(A), matrix(W), jobz='V')
sol = np.array(A[:, DIMS - 1]).reshape(DIMS)
print(np.sum(abs(np.array(psi - old_psi))))
self.sol = sol
self.latent = latent
return sol, latent, threshold
def setUp(self):
from cvxopt import matrix, normal, spdiag, misc, lapack
from ubsdp import ubsdp
m, n = 10, 10
A = normal(m**2, n)
# Z0 random positive definite with maximum e.v. less than 1.0.
Z0 = normal(m, m)
Z0 = Z0 * Z0.T
w = matrix(0.0, (m, 1))
a = +Z0
lapack.syev(a, w, jobz='V')
wmax = max(w)
if wmax > 0.9: w = (0.9 / wmax) * w
Z0 = a * spdiag(w) * a.T
# c = -A'(Z0)
c = matrix(0.0, (n, 1))
misc.sgemv(A,
Z0,
c,
dims={
'l': 0,
'q': [],
's': [m]
},
trans='T',
alpha=-1.0)
# Z1 = I - Z0
Z1 = -Z0
Z1[::m + 1] += 1.0
x0 = normal(n, 1)
X0 = normal(m, m)
X0 = X0 * X0.T
S0 = normal(m, m)
S0 = S0 * S0.T
# B = A(x0) - X0 + S0
B = matrix(A * x0 - X0[:] + S0[:], (m, m))
X = ubsdp(c, A, B)
(self.m, self.n, self.c, self.A, self.B, self.Xubsdp) = (m, n, c, A, B,
X)
def sqrtm(A):
"""Returns the matrix square root of a positive semidefinite matrix."""
if not isinstance(A, (matrix, spmatrix)) or rows(A) != cols(A) or eig(A)[0][0] < 0:
raise TypeError('a symmetric positive semidefinite matrix is required')
V = matrix(A)
z = zeros(rows(A), 1, full=True)
lapack.syev(V, z, jobz='V')
# Round eigenvalues to deal with numerical instability.
# Note: don't use cvxmod atoms pos or sqrt here: overkill.
for i in range(len(z)):
if z[i] <= 0:
z[i] = 0
else:
z[i] = sqrt(z[i])
return V*diag(z)*tp(V)
def fit(self, max_iter=50):
""" Solve the optimization problem with a
sequential convex programming/DC-programming
approach:
Iteratively, find the most likely configuration of
the latent variables and then, optimize for the
model parameter using fixed latent states.
"""
samples = self.sobj.get_num_samples()
dims = self.sobj.get_num_dims()
self.latent = np.random.randint(0, self.sobj.get_num_states(), samples)
self.sol = np.random.randn(dims)
psi = np.zeros((dims, samples))
old_psi = np.zeros((dims, samples))
threshold = 0.
iter = 0
# terminate if objective function value doesn't change much
while iter < max_iter and (iter < 2 or np.sum(np.abs(psi-old_psi)) >= 0.001):
print('Starting iteration {0}.'.format(iter))
print(np.sum(np.abs(psi-old_psi)))
iter += 1
old_psi = psi.copy()
# 1. linearize
# for the current solution compute the
# most likely latent variable configuration
mean = np.zeros(dims)
for i in range(samples):
_, self.latent[i], psi[:, i] = self.sobj.argmax(self.sol, i)
mean += psi[:, i]
mean /= np.float(samples)
mpsi = psi - mean.reshape((dims, 1))
# 2. solve the intermediate convex optimization problem
A = mpsi.dot(mpsi.T)
W = np.zeros((dims, dims))
syev(matrix(A), matrix(W), jobz='V')
self.sol = np.array(A[:, dims-1]).ravel()
return self.sol, self.latent, threshold
def setUp(self):
from cvxopt import matrix, normal, spdiag, misc, lapack
from ubsdp import ubsdp
m, n = 10, 10
A = normal(m**2, n)
# Z0 random positive definite with maximum e.v. less than 1.0.
Z0 = normal(m,m)
Z0 = Z0 * Z0.T
w = matrix(0.0, (m,1))
a = +Z0
lapack.syev(a, w, jobz = 'V')
wmax = max(w)
if wmax > 0.9: w = (0.9/wmax) * w
Z0 = a * spdiag(w) * a.T
# c = -A'(Z0)
c = matrix(0.0, (n,1))
misc.sgemv(A, Z0, c, dims = {'l': 0, 'q': [], 's': [m]}, trans = 'T', alpha = -1.0)
# Z1 = I - Z0
Z1 = -Z0
Z1[::m+1] += 1.0
x0 = normal(n,1)
X0 = normal(m,m)
X0 = X0*X0.T
S0 = normal(m,m)
S0 = S0*S0.T
# B = A(x0) - X0 + S0
B = matrix(A*x0 - X0[:] + S0[:], (m,m))
X = ubsdp(c, A, B)
(self.m, self.n, self.c, self.A, self.B, self.Xubsdp) = (m, n, c, A, B, X)
@author: Nguyen Huu Hiep
'''
from cvxopt import matrix, normal, spdiag, misc, lapack
from ubsdp import ubsdp
m, n = 50, 50
A = normal(m**2, n)
# Z0 random positive definite with maximum e.v. less than 1.0.
Z0 = normal(m, m)
Z0 = Z0 * Z0.T
w = matrix(0.0, (m, 1))
a = +Z0
lapack.syev(a, w, jobz='V')
wmax = max(w)
if wmax > 0.9: w = (0.9 / wmax) * w
Z0 = a * spdiag(w) * a.T
# c = -A'(Z0)
c = matrix(0.0, (n, 1))
misc.sgemv(A, Z0, c, dims={'l': 0, 'q': [], 's': [m]}, trans='T', alpha=-1.0)
# Z1 = I - Z0
Z1 = -Z0
Z1[::m + 1] += 1.0
x0 = normal(n, 1)
X0 = normal(m, m)
X0 = X0 * X0.T
@author: Nguyen Huu Hiep
'''
from cvxopt import matrix, normal, spdiag, misc, lapack
from ubsdp import ubsdp
m, n = 50, 50
A = normal(m**2, n)
# Z0 random positive definite with maximum e.v. less than 1.0.
Z0 = normal(m,m)
Z0 = Z0 * Z0.T
w = matrix(0.0, (m,1))
a = +Z0
lapack.syev(a, w, jobz = 'V')
wmax = max(w)
if wmax > 0.9: w = (0.9/wmax) * w
Z0 = a * spdiag(w) * a.T
# c = -A'(Z0)
c = matrix(0.0, (n,1))
misc.sgemv(A, Z0, c, dims = {'l': 0, 'q': [], 's': [m]}, trans = 'T', alpha = -1.0)
# Z1 = I - Z0
Z1 = -Z0
Z1[::m+1] += 1.0
x0 = normal(n,1)
X0 = normal(m,m)
X0 = X0*X0.T
