def _contour_t(mu, Cov, nu, axes=None, scale=4, transpose=False, colors='k'):
"""
"""
if axes is None:
axes = plt.gca()
if np.shape(mu) != (2,) or np.shape(Cov) != (2,2) or np.shape(nu) != ():
print(np.shape(mu), np.shape(Cov), np.shape(nu))
raise ValueError("Only 2-d t-distribution allowed")
if transpose:
mu = mu[[1,0]]
Cov = Cov[np.ix_([1,0],[1,0])]
s = np.sqrt(np.diag(Cov))
x0 = np.linspace(mu[0]-scale*s[0], mu[0]+scale*s[0], num=100)
x1 = np.linspace(mu[1]-scale*s[1], mu[1]+scale*s[1], num=100)
X0X1 = misc.grid(x0, x1)
Y = X0X1 - mu
L = linalg.chol(Cov)
logdet_Cov = linalg.chol_logdet(L)
Z = linalg.chol_solve(L, Y)
Z = linalg.inner(Y, Z, ndim=1)
lpdf = random.t_logpdf(Z, logdet_Cov, nu, 2)
p = np.exp(lpdf)
shape = (np.size(x0), np.size(x1))
X0 = np.reshape(X0X1[:,0], shape)
X1 = np.reshape(X0X1[:,1], shape)
P = np.reshape(p, shape)
return axes.contour(X0, X1, P, colors=colors)
def gaussian_mixture_logpdf(x, w, mu, Sigma):
# Shape(x) = (N, D)
# Shape(w) = (K,)
# Shape(mu) = (K, D)
# Shape(Sigma) = (K, D, D)
# Shape(result) = (N,)
# Dimensionality
D = np.shape(x)[-1]
# Cholesky decomposition of the covariance matrix
U = linalg.chol(Sigma)
# Reshape x:
# Shape(x) = (N, 1, D)
x = np.expand_dims(x, axis=-2)
# (x-mu) and (x-mu)'*inv(Sigma)*(x-mu):
# Shape(v) = (N, K, D)
# Shape(z) = (N, K)
v = x - mu
z = np.einsum('...i,...i', v, linalg.chol_solve(U, v))
# Log-determinant of Sigma:
# Shape(ldet) = (K,)
ldet = linalg.chol_logdet(U)
# Compute log pdf for each cluster:
# Shape(lpdf) = (N, K)
lpdf = misc.gaussian_logpdf(z, 0, 0, ldet, D)
def check_X(self, mask):
"""
Check update equations of parent X.
Use a simple model that has parent nodes and a child node around
MatrixDot.
"""
# Parent nodes
A = Gaussian(self.a.reshape((self.D3,self.D2,1,self.M*self.N)),
np.identity(self.M*self.N),
name='A')
X = Gaussian(np.zeros(self.N),
np.identity(self.N),
plates=(self.D3,1,self.D1),
name='X')
# Node itself
AX = MatrixDot(A, X, name='AX')
# Child node
Y = Gaussian(AX,
self.Lambda,
plates=(self.D4,self.D3,self.D2,self.D1),
name='Y')
# Put in data
Y.observe(self.y, mask=mask)
# VB model
Q = VB(Y, X)
Q.update(X, repeat=1)
u = X.get_moments()
# Compute true solution
mask = np.ones((self.D4,self.D3,self.D2,self.D1)) * mask
vv = mask[...,np.newaxis,np.newaxis] * self.vv
aa = np.einsum('...ki,...kl,...lj->...ij', self.a, vv, self.a)
aa = aa + np.einsum('...ij,...kl,...kl->...ij',
np.identity(self.N),
np.identity(self.M),
vv)
aa = np.sum(aa, axis=2, keepdims=True)
aa = np.sum(aa, axis=0)
Cov_X = linalg.chol_inv(linalg.chol(aa + np.identity(self.N)))
mu_X = np.einsum('...kj,...kl,...l->...j',
self.a,
vv,
self.y)
mu_X = np.sum(mu_X, axis=2, keepdims=True)
mu_X = np.sum(mu_X, axis=0)
mu_X = np.einsum('...ij,...j->...i', Cov_X, mu_X)
# Compare VB results to the analytic solution:
Cov_vb = u[1] - u[0][...,np.newaxis,:]*u[0][...,:,np.newaxis]
testing.assert_allclose(Cov_vb, Cov_X,
err_msg="Incorrect second moment.")
testing.assert_allclose(u[0], mu_X,
err_msg="Incorrect first moment.")
def check_A(self, mask):
"""
Check update equations of parent A.
Use a simple model that has parent nodes and a child node around
MatrixDot.
"""
# Parent nodes
A = Gaussian(np.zeros(self.M*self.N),
np.identity(self.M*self.N),
plates=(self.D3,self.D2,1),
name='A')
X = Gaussian(self.x, np.identity(self.N), name='X')
# Node itself
AX = MatrixDot(A, X, name='AX')
# Child node
Y = Gaussian(AX,
self.Lambda,
plates=(self.D4,self.D3,self.D2,self.D1),
name='Y')
# Put in data
Y.observe(self.y, mask=mask)
# VB model
Q = VB(Y, A)
Q.update(A, repeat=1)
u = A.get_moments()
# Compute true solution
mask = np.ones((self.D4,self.D3,self.D2,self.D1)) * mask
vv = mask[...,np.newaxis,np.newaxis] * self.vv
xx = (self.x[...,:,np.newaxis] * self.x[...,np.newaxis,:]
+ np.identity(self.N))
Cov_A = (vv[...,:,np.newaxis,:,np.newaxis] *
xx[...,np.newaxis,:,np.newaxis,:]) # from data
Cov_A = np.sum(Cov_A, axis=(0,3), keepdims=True)
Cov_A = np.reshape(Cov_A,
(self.D3,self.D2,1,self.M*self.N,self.M*self.N))
Cov_A = Cov_A + np.identity(self.M*self.N) # add prior
Cov_A = linalg.chol_inv(linalg.chol(Cov_A))
mu_A = np.einsum('...ij,...j,...k->...ik', vv, self.y, self.x)
mu_A = np.sum(mu_A, axis=(0,3))
mu_A = np.reshape(mu_A, (self.D3,self.D2,1,self.M*self.N))
mu_A = np.einsum('...ij,...j->...i', Cov_A, mu_A)
# Compare VB results to the analytic solution:
Cov_vb = u[1] - u[0][...,np.newaxis,:]*u[0][...,:,np.newaxis]
testing.assert_allclose(Cov_vb, Cov_A,
err_msg="Incorrect second moment.")
testing.assert_allclose(u[0], mu_A,
err_msg="Incorrect first moment.")
def compute_fixed_moments(self, Lambda, gradient=None):
""" Compute moments for fixed x. """
L = linalg.chol(Lambda, ndim=self.ndim)
ldet = linalg.chol_logdet(L, ndim=self.ndim)
u = [Lambda,
ldet]
if gradient is None:
return u
du0 = gradient[0]
du1 = (
misc.add_trailing_axes(gradient[1], 2*self.ndim)
* linalg.chol_inv(L, ndim=self.ndim)
)
du = du0 + du1
return (u, du)
def compute_fixed_moments_and_f(self, Lambda, mask=True):
r"""
Compute u(x) and f(x) for given x.
.. math:
u(\Lambda) =
\begin{bmatrix}
\Lambda
\\
\log |\Lambda|
\end{bmatrix}
"""
k = np.shape(Lambda)[-1]
ldet = linalg.chol_logdet(linalg.chol(Lambda))
u = [Lambda,
ldet]
f = -(k+1)/2 * ldet
return (u, f)
def compute_moments_and_cgf(self, phi, mask=True):
r"""
Return moments and cgf for given natural parameters
.. math::
\langle u \rangle =
\begin{bmatrix}
\phi_2 (-\phi_1)^{-1}
\\
-\log|-\phi_1| + \psi_k(\phi_2)
\end{bmatrix}
\\
g(\phi) = \phi_2 \log|-\phi_1| - \log \Gamma_k(\phi_2)
"""
U = linalg.chol(-phi[0])
k = np.shape(phi[0])[-1]
#k = self.dims[0][0]
logdet_phi0 = linalg.chol_logdet(U)
u0 = phi[1][...,np.newaxis,np.newaxis] * linalg.chol_inv(U)
u1 = -logdet_phi0 + misc.multidigamma(phi[1], k)
u = [u0, u1]
g = phi[1] * logdet_phi0 - special.multigammaln(phi[1], k)
return (u, g)
def kalman_filter(y, U, A, V, mu0, Cov0, out=None):
"""
Perform Kalman filtering to obtain filtered mean and covariance.
The parameters of the process may vary in time, thus they are
given as iterators instead of fixed values.
Parameters
----------
y : (N,D) array
"Normalized" noisy observations of the states, that is, the
observations multiplied by the precision matrix U (and possibly
other transformation matrices).
U : (N,D,D) array or N-list of (D,D) arrays
Precision matrix (i.e., inverse covariance matrix) of the observation
noise for each time instance.
A : (N-1,D,D) array or (N-1)-list of (D,D) arrays
Dynamic matrix for each time instance.
V : (N-1,D,D) array or (N-1)-list of (D,D) arrays
Covariance matrix of the innovation noise for each time instance.
Returns
-------
mu : array
Filtered mean of the states.
Cov : array
Filtered covariance of the states.
See also
--------
rts_smoother
"""
mu = mu0
Cov = Cov0
# Allocate memory for the results
(N, D) = np.shape(y)
X = np.empty((N, D))
CovX = np.empty((N, D, D))
# Update step for t=0
M = np.dot(np.dot(Cov, U[0]), Cov) + Cov
L = linalg.chol(M)
mu = np.dot(Cov, linalg.chol_solve(L, np.dot(Cov, y[0]) + mu))
Cov = np.dot(Cov, linalg.chol_solve(L, Cov))
X[0, :] = mu
CovX[0, :, :] = Cov
# for (yn, Un, An, Vn) in zip(y, U, A, V):
for n in range(len(y) - 1): # (yn, Un, An, Vn) in zip(y, U, A, V):
# Prediction step
mu = np.dot(A[n], mu)
Cov = np.dot(np.dot(A[n], Cov), A[n].T) + V[n]
# Update step
M = np.dot(np.dot(Cov, U[n + 1]), Cov) + Cov
L = linalg.chol(M)
mu = np.dot(Cov, linalg.chol_solve(L, np.dot(Cov, y[n + 1]) + mu))
Cov = np.dot(Cov, linalg.chol_solve(L, Cov))
# Force symmetric covariance (for numeric inaccuracy)
Cov = 0.5 * Cov + 0.5 * Cov.T
# Store results
X[n + 1, :] = mu
CovX[n + 1, :, :] = Cov
return (X, CovX)
def compute_fixed_moments(self, Lambda):
""" Compute moments for fixed x. """
ldet = linalg.chol_logdet(linalg.chol(Lambda))
u = [Lambda,
ldet]
return u
def compute_fixed_moments(self, Lambda):
""" Compute moments for fixed x. """
ldet = linalg.chol_logdet(linalg.chol(Lambda))
u = [Lambda, ldet]
return u
