def star(SA, SB, device='cpu'):
'''
STAR Redheffer Star Product
% INPUT ARGUMENTS
% ================
% SA First Scattering Matrix
% .S11 S11 scattering parameter
% .S12 S12 scattering parameter
% .S21 S21 scattering parameter
% .S22 S22 scattering parameter
%
% SB Second Scattering Matrix
% .S11 S11 scattering parameter
% .S12 S12 scattering parameter
% .S21 S21 scattering parameter
% .S22 S22 scattering parameter
%
% OUTPUT ARGUMENTS
% ================
% S Combined Scattering Matrix
% .S11 S11 scattering parameter
% .S12 S12 scattering parameter
% .S21 S21 scattering parameter
% .S22 S22 scattering parameter
'''
if device == 'gpu' or device == 'cuda': # use GPU, cupy
N = SA['S11'].shape[0]
I = cp.eye(N)
S11 = SA['S11'] + (SA['S12'] @ CLA.inv(I - (SB['S11'] @ SA['S22']))
@ SB['S11'] @ SA['S21'])
S12 = SA['S12'] @ CLA.inv(I - (SB['S11'] @ SA['S22'])) @ SB['S12']
S21 = SB['S21'] @ CLA.inv(I - (SA['S22'] @ SB['S11'])) @ SA['S21']
S22 = SB['S22'] + (SB['S21'] @ CLA.inv(I - (SA['S22'] @ SB['S11']))
@ SA['S22'] @ SB['S12'])
S = {'S11': S11, 'S12': S12, 'S21': S21, 'S22': S22}
else: # cpu, numpy
N = SA['S11'].shape[0]
I = np.eye(N)
S11 = SA['S11'] + (SA['S12'] @ LA.inv(I - (SB['S11'] @ SA['S22']))
@ SB['S11'] @ SA['S21'])
S12 = SA['S12'] @ LA.inv(I - (SB['S11'] @ SA['S22'])) @ SB['S12']
S21 = SB['S21'] @ LA.inv(I - (SA['S22'] @ SB['S11'])) @ SA['S21']
S22 = SB['S22'] + (SB['S21'] @ LA.inv(I - (SA['S22'] @ SB['S11']))
@ SA['S22'] @ SB['S12'])
S = {'S11': S11, 'S12': S12, 'S21': S21, 'S22': S22}
return S
def _eigh(A, B=None):
"""
Helper function for converting a generalized eigenvalue problem
A(X) = lambda(B(X)) to standard eigen value problem using cholesky
transformation
"""
if (B is None): # use cupy's eigh in standard case
vals, vecs = linalg.eigh(A)
return vals, vecs
R = _cholesky(B)
RTi = linalg.inv(R)
Ri = linalg.inv(R.T)
F = cupy.matmul(RTi, cupy.matmul(A, Ri))
vals, vecs = linalg.eigh(F)
eigVec = cupy.matmul(Ri, vecs)
return vals, eigVec
def _b_orthonormalize(B, blockVectorV, blockVectorBV=None, retInvR=False):
"""B-orthonormalize the given block vector using Cholesky."""
normalization = blockVectorV.max(axis=0) + cupy.finfo(
blockVectorV.dtype).eps
blockVectorV = blockVectorV / normalization
if blockVectorBV is None:
if B is not None:
blockVectorBV = B(blockVectorV)
else:
blockVectorBV = blockVectorV
else:
blockVectorBV = blockVectorBV / normalization
VBV = cupy.matmul(blockVectorV.T.conj(), blockVectorBV)
try:
# VBV is a Cholesky factor
VBV = _cholesky(VBV)
VBV = linalg.inv(VBV.T)
blockVectorV = cupy.matmul(blockVectorV, VBV)
if B is not None:
blockVectorBV = cupy.matmul(blockVectorBV, VBV)
else:
blockVectorBV = None
except numpy.linalg.LinAlgError:
# LinAlg Error: cholesky transformation might fail in rare cases
# raise ValueError("cholesky has failed")
blockVectorV = None
blockVectorBV = None
VBV = None
if retInvR:
return blockVectorV, blockVectorBV, VBV, normalization
else:
return blockVectorV, blockVectorBV
def loggrad_z(thetas):
"""
theta = [z, sigma_z_sqr, w, gamma, eta]
"""
z = thetas['z']
w = thetas['w']
gamma = thetas['gamma']
eta = thetas['eta']
sigma_z_sqr = thetas['sigma_z_sqr']
K = z.shape[0]
# Precision matrix with covariance [1, 1.98; 1.98, 4].
# A = np.linalg.inv( cov )
V_eta = np.exp(np.dot(w, gamma))
mu_eta = np.dot(w,z)
V_z = inv(np.dot(w.T,np.dot(np.diag(1/V_eta), w)) + 1/sigma_z_sqr * np.diag(np.ones(K)))
mu_z = np.dot(V_z, np.dot(w.T,np.dot(np.diag(1/V_eta), eta)))
logp = -0.5 * np.dot((z - mu_z).T, np.dot(inv(V_z), z-mu_z))-np.log(norm.cdf(mu_eta/(V_eta**0.5))).sum()
grad = - np.dot(inv(V_z), z) + np.dot(inv(V_z), mu_z) - np.dot(w.T,norm.pdf(mu_eta/(V_eta**0.5))/(norm.cdf(mu_eta/(V_eta**0.5)) * V_eta ** 0.5))
return -logp, -grad
def loggrad_delta(thetas):
"""
theta = [delta, sigma_delta_sqr, pi, u, xi]
"""
pi = thetas['pi']
xi = thetas['xi']
delta = thetas['delta']
u = thetas['u']
sigma_delta_sqr = thetas['sigma_delta_sqr']
K = delta.shape[0]
# Precision matrix with covariance [1, 1.98; 1.98, 4].
# A = np.linalg.inv( cov )
V_u = np.exp(np.dot(pi, xi))
mu_u = np.dot(pi,delta)
V_delta = inv(np.dot(pi.T,np.dot(np.diag(1/V_u), pi)) + 1/sigma_delta_sqr * np.diag(np.ones(K)))
mu_delta = np.dot(V_delta, np.dot(pi.T,np.dot(np.diag(1/V_u), u)))
logp = -0.5 * np.dot((delta - mu_delta).T, np.dot(inv(V_delta), delta-mu_delta))-np.log(norm.cdf(mu_u/(V_u**0.5))).sum()
grad = - np.dot(inv(V_delta), delta) + np.dot(inv(V_delta), mu_delta) - np.dot(pi.T,norm.pdf(mu_u/(V_u**0.5))/(norm.cdf(mu_u/(V_u**0.5)) * V_u ** 0.5))
return -logp, -grad
def Correct(self, xHatMinus, PMinus, z, active_measurements=None):
'''For use in Iterate'''
if active_measurements is not None:
H = self.H * active_measurements
else:
H = self.H
K = dot(PMinus, dot(H.T,
linalg.inv(dot(H, dot(PMinus, H.T)) + self.R)))
xHat = xHatMinus + dot(K, (z - dot(H, xHatMinus)))
P = dot((self.I - dot(K, H)), PMinus)
return K, xHat, P
def update(self, z):
# Transform sigmas into measurement space
for i in range(self.num_sp):
self.sp_h[i] = self.H(self.sp_f[i])
# Mean and covariance of the state in the measurement space
Hx_bar, PHx = unscented_transform(self.sp_h, self.weights, self.R)
# Cross variance of Fx and Hx -- used to calculate K
cross_var = np.zeros((self.Nx, self.Nz))
for i in range(self.num_sp):
cross_var += self.weights[i] * np.outer(self.sp_f[i] - self.x_hat,
self.sp_h[i] - Hx_bar)
# Calculate K and measurement residual
K = np.dot(cross_var, inv(PHx))
residual = self.difference(np.array(z), Hx_bar)
# Update predicted to new values for state and variance
self.x_hat += np.dot(K, residual)
self.P -= np.dot(K, np.dot(PHx, K.T))
def inv(self, a):
return CuPyTensor(la.inv(a.unwrap()))
def basis_pursuit(A, b, tol=1e-4, niter=100, biter=32):
"""
solves min |x|_1 s.t. Ax=b using a Primal-Dual Interior Point Method
Args:
A: design matrix of size (d, n)
b: measurement vector of length d
tol: solver tolerance
niter: maximum length of central path
biter: maximum number of steps in backtracking line search
Returns:
vector of length n
"""
A = cp.asarray(A)
b = cp.asarray(b)
d, n = A.shape
alpha = 0.01
beta = 0.5
mu = 10
e = cp.ones(n)
gradf0 = cp.hstack([cp.zeros(n), e])
x = (A.T).dot(inv(A.dot(A.T))).dot(b)
absx = cp.abs(x)
u = 0.95 * absx + 0.1 * cp.max(absx)
fu1 = x - u
fu2 = -x - u
lamu1 = -1.0 / fu1
lamu2 = -1.0 / fu2
v = A.dot(lamu2 - lamu1)
ATv = (A.T).dot(v)
sdg = -(cp.inner(fu1, lamu1) + cp.inner(fu2, lamu2))
tau = 2.0 * n * mu / sdg
ootau = 1.0 / tau
rcent = cp.hstack([-lamu1 * fu1, -lamu2 * fu2]) - ootau
rdual = gradf0 + cp.hstack([lamu1 - lamu2 + ATv, -lamu1 - lamu2])
rpri = A.dot(x) - b
resnorm = cp.sqrt(norm(rdual)**2 + norm(rcent)**2 + norm(rpri)**2)
rdp = cp.empty(2 * n)
rcp = cp.empty(2 * n)
for i in range(niter):
oofu1 = 1.0 / fu1
oofu2 = 1.0 / fu2
w1 = -ootau * (oofu2 - oofu1) - ATv
w2 = -1.0 - ootau * (oofu1 + oofu2)
w3 = -rpri
lamu1xoofu1 = lamu1 * oofu1
lamu2xoofu2 = lamu2 * oofu2
sig1 = -lamu1xoofu1 - lamu2xoofu2
sig2 = lamu1xoofu1 - lamu2xoofu2
sigx = sig1 - sig2**2 / sig1
if cp.min(cp.abs(sigx)) == 0.0:
break
w1p = -(w3 - A.dot(w1 / sigx - w2 * sig2 / (sigx * sig1)))
H11p = A.dot((A.T) * (e / sigx)[:, cp.newaxis])
if cp.min(sigx) > 0.0:
dv = solve(H11p, w1p)
else:
dv = solve(H11p, w1p)
dx = (w1 - w2 * sig2 / sig1 - (A.T).dot(dv)) / sigx
Adx = A.dot(dx)
ATdv = (A.T).dot(dv)
du = (w2 - sig2 * dx) / sig1
dlamu1 = lamu1xoofu1 * (du - dx) - lamu1 - ootau * oofu1
dlamu2 = lamu2xoofu2 * (dx + du) - lamu2 - ootau * oofu2
s = 1.0
indp = cp.less(dlamu1, 0.0)
indn = cp.less(dlamu2, 0.0)
if cp.any(indp):
s = min(s, cp.min(-lamu1[indp] / dlamu1[indp]))
if cp.any(indn):
s = min(s, cp.min(-lamu2[indn] / dlamu2[indn]))
indp = cp.greater(dx - du, 0.0)
indn = cp.greater(-dx - du, 0.0)
if cp.any(indp):
s = min(s, cp.min(-fu1[indp] / (dx[indp] - du[indp])))
if cp.any(indn):
s = min(s, cp.min(-fu2[indn] / (-dx[indn] - du[indn])))
s = 0.99 * s
for j in range(biter):
xp = x + s * dx
up = u + s * du
vp = v + s * dv
ATvp = ATv + s * ATdv
lamu1p = lamu1 + s * dlamu1
lamu2p = lamu2 + s * dlamu2
fu1p = xp - up
fu2p = -xp - up
rdp[:n] = lamu1p - lamu2p + ATvp
rdp[n:] = -lamu1p - lamu2p
rdp += gradf0
rcp[:n] = -lamu1p * fu1p
rcp[n:] = lamu2p * fu2p
rcp -= ootau
rpp = rpri + s * Adx
s *= beta
if (cp.sqrt(norm(rdp)**2 + norm(rcp)**2 + norm(rpp)**2) <=
(1 - alpha * s) * resnorm):
break
else:
break
x = xp
lamu1 = lamu1p
lamu2 = lamu2p
fu1 = fu1p
fu2 = fu2p
sdg = -(cp.inner(fu1, lamu1) + cp.inner(fu2, lamu2))
if sdg < tol:
return cp.asnumpy(x)
u = up
v = vp
ATv = ATvp
tau = 2.0 * n * mu / sdg
rpri = rpp
rcent[:n] = lamu1 * fu1
rcent[n:] = lamu2 * fu2
ootau = 1.0 / tau
rcent -= ootau
rdual[:n] = lamu1 - lamu2 + ATv
rdual[n:] = -lamu1 + lamu2
rdual += gradf0
resnorm = cp.sqrt(norm(rdual)**2 + norm(rcent)**2 + norm(rpri)**2)
return cp.asnumpy(x)
accpt_num_xi = 0
accpt_num_z = 0
accpt_num_gamma = 0
S = 11000
all_beta = np.zeros([S, K])
all_xi = np.zeros([S,K])
all_delta = np.zeros([S,K])
all_z = np.zeros([S,K])
all_gamma = np.zeros([S,K])
all_sigma_v_sqr = np.zeros([S,])
all_sigma_alpha_sqr = np.zeros([S,])
for i in range(S):
print(i)
### Posterior
# beta
V_beta = inv((np.dot(x.T,x) * sigma_beta_sqr + sigma_v_sqr)/(sigma_v_sqr * sigma_beta_sqr))
mu_beta = np.dot(V_beta, np.dot(x.T, y-u-np.kron(alpha, np.ones([T,])) - np.kron(eta, np.ones(T,)))/sigma_v_sqr)
beta = multivariate_normal.rvs(mu_beta, V_beta)
# sigma_v_sqr
y_tilda = y-np.dot(x,beta)-u-np.kron(alpha,np.ones(T,))-np.kron(eta,np.ones(T,))
shape = (NT+1)/2
scale = 2 / (0.0001 + np.dot(y_tilda.T, y_tilda))
sigma_v_sqr = 1/np.random.gamma(shape,scale)
# sigma_alpha_sqr
shape = (N+1)/2
scale = 2/ (0.0001 + np.dot(alpha.T,alpha))
sigma_alpha_sqr = 1/np.random.gamma(shape,scale)
# alpha
# scale = sigma_alpha_sqr * sigma_v_sqr / (T * sigma_alpha_sqr + sigma_v_sqr)
# y_bar = y-np.dot(x, beta)-np.kron(eta, np.ones(T,))-u
