def main():
n=10
# prepare matrix a
a=np.diag([5.]*n)
a+=np.diagflat([2.]*(n-1),1)
a+=np.diagflat([2.]*(n-1),-1)
print(a)
b=np.array([3,1,4,0,5,-1,6,-2,7,-15],dtype='f').T
#initial value x
x=np.ones(10).T
D=np.diag(np.diag(a))
L=np.tril(a,-1)
U=np.triu(a,+1)
M= -np.linalg.inv(D) @ (L+U)
N=np.linalg.inv(D)
for k in range(maxiteration):
x_new=M @ x + N @ b
if(np.linalg.norm(x_new-x) <epsilon):
break
x=x_new
else:
print("fail jacobi method ...")
exit(1)
print("the sol of ax = b using Jacobi method is \n{}".format(x))
print("iteration {} times".format(k))
print("indeed ax-b is \n{}".format(a @x -b))
print("you can check sol x using sympy...")
sy_a=sy.Matrix(a)
sy_x=sy_a.solve(b)
print(sy_x)
def matrix_simulstionnp(self, temp, dt, t, outtemps):
size = len(self.npwam[0])
time = 0
dt = dt
I = np.diagflat(np.array([[1]*size]))
neg_c = np.diagflat((self.npwam.dot(np.array([[-1]]*size))))
M = self.npthermal_mass.dot(self.npwam + neg_c).dot(dt) + I
print M, "Tdot"
temp = np.array(temp) # K
#temps = [x for x in temp.matrix]
temps = [[cell for cell in row] for row in temp]
# finish matrix simulation
hour = 60*60*3
idx = 0
qhour = 15*60
while time < t:
if hour == (60*60*3):
temp[3][0] = outtemps[idx][0]
hour = 0
idx += 1
tempn = M.dot(temp)
temp = tempn
if qhour >= 15*60:
for i in range(len(tempn)):
temps[i].append(temp[i][0])
qhour = 0
time += dt
hour += dt
qhour += dt
return temps
def rvmtrain(xinp, yinp, r, beta = 100):
dmat = getDesMat(xinp, r)
N = len(yinp)
target = yinp
alphas = np.ones((len(xinp) +1, 1))
Amat = np.diagflat(alphas)
newAlphas = np.copy(alphas)
converged = False
idx = np.ones(len(alphas)) == 1
mMat = np.zeros(len(alphas))
iterationCount = 0
for t in range(5000):
iterationCount = iterationCount + 1
idx = np.abs(newAlphas) < alphaThreshold
idx = np.squeeze(idx)
sig = Amat[idx][:,idx] + beta * np.dot(dmat[:,idx].transpose(), dmat[:,idx])
sigMat = np.linalg.inv(sig)
mMat[idx] = beta * np.dot(sigMat, np.dot(dmat[:,idx].transpose(), target))
oldAlphas = np.copy(newAlphas)
gamma = 1 - np.transpose(newAlphas[idx]) * np.diag(sigMat)
newAlphas[idx] = np.transpose( gamma / np.array(map(float,mMat[idx]**2)) )
beta = ( N - np.sum(gamma) ) / np.linalg.norm( yinp - np.dot(dmat[:,idx], mMat[idx]) )
delta = sum(np.abs(newAlphas - oldAlphas))
# if (sum(newAlphas<0)>0):
# print iterationCount
if (delta < convThreshold):
print "\n\n\n\n\n!!!!!CONVERGED!!!!!\n\n\n\n\n"
converged = True
print newAlphas
break
Amat = np.diagflat(newAlphas)
relevant_vecs_ind = []
x_rel =[]
y_rel = []
print "iterations: {}".format(iterationCount)
# If we start from 0 then we check if the bias term of alpha is relevant. If it is so we pick the last training point
# x[-1] and y[-1] to be a relevancevector. The bias term is problematic.
for i in range(1,N+1):
if newAlphas[i] < alphaThreshold:
relevant_vecs_ind.append(i+1)
x_rel.append(xinp[i-1])
y_rel.append(yinp[i-1])
print "number of relevancevectors (alpha < {0}): ".format(alphaThreshold) + str(np.sum(idx[1:]))
muMat = mMat
print "beta: " + str(beta)
# for a in newAlphas:
# print a
return muMat, beta, converged, idx, x_rel, y_rel
def MStep(self, posterior, data, mix_pi=None):
if isinstance(data, DataSet):
x = data.internalData
elif hasattr(data, "__iter__"):
x = data
else:
raise TypeError, "Unknown/Invalid input to MStep."
post = posterior.sum() # sum of posteriors
self.mean = np.dot(posterior, x) / post
# centered input values (with new mus)
centered = np.subtract(x, np.repeat([self.mean], len(x), axis=0));
# estimating correlation factor
sigma = np.dot(np.transpose(np.dot(np.identity(len(posterior)) * posterior, centered)), centered) / post # sigma/covariance matrix
diagsigma = np.diagflat(1.0 / np.diagonal(sigma)) # vector with diagonal entries of sigma matrix
correlation = np.dot(np.dot(diagsigma, np.multiply(sigma, sigma)), diagsigma) # correlation matrix with entries sigma_xy^2/(sigma^2_x * sigma^2_y)
correlation = correlation - np.diagflat(np.diagonal(correlation)) # making diagonal entries = 0
# XXX - check this
parents = self.maximunSpanningTree(correlation) # return maximun spanning tree from the correlation matrix
self.parents = self.directTree(parents, 0) # by default direct tree from 0
# XXX note that computational time could be saved as these functions share same suficient statistics
ConditionalGaussDistribution.MStep(self, posterior, data, mix_pi)
def ARD(X,Y):
# X - (p x q) matrix with inputs in rows
# Y - (p, 1) matrix with measurements
# Implelements the ARD regression, adapted from:
#M. Sahani and J. F. Linden.
#Evidence optimization techniques for estimating stimulus-response functions.
#In S. Becker, S. Thrun, and K. Obermayer, eds., Advances in Neural Information Processing Systems, vol. 15, pp. 301-308, Cambridge, MA, 2003.
(p,q) = numpy.shape(X)
#initialize parameters
sigma_sq = 0.1
CC = X.T * X
XY = X.T * Y
start_flag = False
alpha = numpy.mat(numpy.zeros((q,1)))+2.0
for i in xrange(0,100):
sigma = numpy.linalg.inv(CC/sigma_sq + numpy.diagflat(alpha))
ni = sigma * (XY) / (sigma_sq)
sigma_sq = numpy.sum(numpy.power(Y - X*ni,2))/(p - numpy.sum(1 - numpy.multiply(numpy.mat(numpy.diagonal(sigma)).T,alpha)));
print numpy.min(numpy.abs(ni))
alpha = numpy.mat(numpy.divide((1 - numpy.multiply(alpha,numpy.mat(numpy.diagonal(sigma)).T)) , numpy.power(ni,2)))
print sigma_sq
w = numpy.linalg.inv(CC + sigma_sq * numpy.diagflat(alpha)) * (XY)
print alpha
print sigma_sq
return w
def burgers_ei_bcp(M, N, xmin, xmax, tinz, tfin, ic):
dx = (xmax - xmin) / M
dt = (tfin - tinz) / N
x = xmin + array(range(0, M + 1)) * dx
t = array(range(0, N + 1)) * dt
u = zeros([M + 1, N + 1])
u[:, 0] = ic(x)
uold = zeros([M + 1])
uold[0:M + 1] = u[:, 0]
for n in range(0, N + 1):
# numero di Courant
nu = max(abs(uold)) * dt / dx
# CFL
cfl = nu * dt / dx
if cfl > 1:
print "CFL " + str(cfl)
raise ArithmeticError
else:
A = eye(M + 1) + 0.5 * dt * diagflat(uold[0:M], 1) / dx - 0.5 * dt * diagflat(uold[1:M + 1], -1) / dx
A[0, M] = -0.5 * dt * uold[0] / dx
A[M, 1] = 0.5 * dt * uold[M] / dx
unew = solve(A, uold.transpose()).transpose()
uold = unew
u[0:M + 1, n] = unew
return u, x, t
def gen_roots_and_weights(n, an_func, sqrt_bn_func, mu):
"""[x,w] = gen_roots_and_weights(n,an_func,sqrt_bn_func,mu)
Returns the roots (x) of an nth order orthogonal polynomial,
and weights (w) to use in appropriate Gaussian quadrature with that
orthogonal polynomial.
The polynomials have the recurrence relation
P_n+1(x) = (x - A_n) P_n(x) - B_n P_n-1(x)
an_func(n) should return A_n
sqrt_bn_func(n) should return sqrt(B_n)
mu ( = h_0 ) is the integral of the weight over the orthogonal interval
"""
nn = np.arange(1.0,n)
sqrt_bn = sqrt_bn_func(nn)
an = an_func(np.concatenate(([0], nn)))
x, v = eig((np.diagflat(an) +
np.diagflat(sqrt_bn,1) +
np.diagflat(sqrt_bn,-1)))
answer = []
sortind = x.real.argsort()
answer.append(x[sortind])
answer.append((mu*v[0]**2)[sortind])
return answer
def trasporto_ei_bcp(M, N, xmin, xmax, tinz, tfin, vel, ic):
dx = (xmax - xmin) / M
dt = (tfin - tinz) / N
x = xmin + array(range(0, M + 1)) * dx
t = array(range(0, N + 1)) * dt
u = zeros([M + 1, N + 1])
u[:, 0] = ic(x)
uold = zeros([M + 1])
uold[0:M + 1] = u[:, 0]
a = vel
# CFL
cfl = a * dt / dx
# numero di Courant
nu = a * dt / dx
A = eye(M + 1) + 0.5 * nu * diagflat(ones([1, M]), 1) - 0.5 * nu * diagflat(ones([1, M]), -1)
A[0, M] = -0.5 * nu
A[M, 1] = 0.5 * nu
if cfl > 1:
print "CFL " + str(cfl)
raise ArithmeticError
else:
for n in range(0, N + 1):
unew = solve(A, uold.transpose()).transpose()
uold = unew
u[0:M + 1, n] = unew
return u, x, t
def fit_map(self, X, y, full_posterior=True):
"""
Fit a MAP estimate
X: features, n_samples by n_features nd-array
y: target values, n_samples array
"""
if self.fit_intercept:
X = self.add_intercept(X)
#data setup
f_dim = X.shape[1]
if np.isscalar(self.inv_lamb):
inv_lamb = np.diagflat(np.repeat(self.inv_lamb,f_dim))
else:
inv_lamb = np.diagflat(self.inv_lamb)
if np.isscalar(self.beta_mu):
beta_mu = np.repeat(self.beta_mu,f_dim)
else:
beta_mu = self.beta_mu
sigma = self.sigma
#let the actual calculation begin
l = sigma**2 * inv_lamb
s = LA.inv(X.T.dot(X)+l)
# adding in the mean of the prior
b0 = sigma**2 * inv_lamb.dot(beta_mu)
self.beta = s.dot(X.T.dot(y) + b0)
if full_posterior:
self.Sigma = sigma * sigma * s
def jac_dtwa(s, t, param):
"""
Jacobian of the general case. First order.
This is given by 9 NXN submatrices:
J00=J11=J22=0
Although Jacobian is NOT antisymmetric in general! See below
J01 = +J_z diag(J|s^x>) + h(t) h_z - J_y (J#|s^z>)
J10 = -J_z diag(J|s^x>) - h(t) h_z + J_x (J#|s^z>)
J02 = -J_y diag(J|s^y>) - h(t) h_y + J_z (J#|s^y>)
J20 = +J_y diag(J|s^y>) + h(t) h_y - J_x (J#|s^y>)
J12 = +J_x diag(J|s^x>) + h(t) h_x - J_z (J#|s^x>)
J21 = -J_x diag(J|s^x>) - h(t) h_x + J_y (J#|s^x>)
Here, '#' (hash operator) means multiply each row of a matrix by the
corresponding vector element. This is implemented by numpy.multiply()
"""
N = param.latsize
# s[0:N] = sx , s[N:2*N] = sy, s[2*N:3*N] = sz
full_jacobian = np.zeros(shape=(3 * N, 3 * N))
diag_jsx = np.diagflat((param.jmat.dot(s[0:N]))) / param.norm
diag_jsy = np.diagflat((param.jmat.dot(s[N : 2 * N]))) / param.norm
# diag_jsz = np.diagflat((param.jmat.dot(s[2*N:3*N])))/param.norm
hash_jsx = (np.multiply(param.jmat.T, s[0:N]).T) / param.norm
hash_jsy = (np.multiply(param.jmat.T, s[N : 2 * N]).T) / param.norm
hash_jsz = (np.multiply(param.jmat.T, s[2 * N : 3 * N]).T) / param.norm
full_jacobian[0:N, N : 2 * N] = param.jz * diag_jsx + drivemat * param.hz - param.jy * hash_jsz
full_jacobian[N : 2 * N, 0:N] = -param.jz * diag_jsx - param.hz + param.jx * hash_jsz
full_jacobian[0:N, 2 * N : 3 * N] = -param.jy * diag_jsy - param.hy + param.jz * hash_jsy
full_jacobian[2 * N : 3 * N, 0:N] = param.jy * diag_jsy + param.hy - param.jx * hash_jsy
full_jacobian[N : 2 * N, 2 * N : 3 * N] = param.jx * diag_jsx + param.hx - param.jz * hash_jsx
full_jacobian[2 * N : 3 * N, N : 2 * N] = -param.jx * diag_jsx - param.hx + param.jy * hash_jsx
return full_jacobian
def get_Hatoms(ph_count, at_count, wc, wa, g): #------------------------------------------------------------------------------------------------------------------ sigmadiag = [1] sigmacross = np.diagflat(sigmadiag, -1) sigma = np.diagflat(sigmadiag, 1) sigmacrosssigma = np.dot(sigmacross, sigma) #------------------------------------------------------------------------------------------------------------------ ph_dim = ph_count+1 I_ph = np.identity(ph_dim) #------------------------------------------------------------------------------------------------------------------ H_dim = (ph_count+1) * pow(2, at_count) H_atoms = np.zeros([H_dim, H_dim]) #------------------------------------------------------------------------------------------------------------------ for i in range(1, at_count+1): elem = sigmacrosssigma at_prev = np.identity(pow(2, i-1)) elem = np.kron(at_prev, elem) at_next = np.identity( pow(2, at_count-i)) elem = np.kron(elem, at_next) H_atoms += wa * np.kron(I_ph, elem) #------------------------------------------------------------------------------------------------------------------ return H_atoms
def get_Hint_EXACT(ph_count, at_count, wc, wa, g): #------------------------------------------------------------------------------------------------------------------ adiag = np.sqrt(np.arange(1, ph_count+1)) across = np.diagflat(adiag, -1) a = np.diagflat(adiag, 1) acrossa = np.dot(across, a) #------------------------------------------------------------------------------------------------------------------ sigmadiag = [1] sigmacross = np.diagflat(sigmadiag, -1) sigma = np.diagflat(sigmadiag, 1) sigmacrosssigma = np.dot(sigmacross, sigma) #------------------------------------------------------------------------------------------------------------------ H_dim = (ph_count+1) * pow(2, at_count) H_int = np.zeros([H_dim, H_dim], dtype=complex) #------------------------------------------------------------------------------------------------------------------ for i in range(1, at_count+1): #------------------------------------------------ elem = (across + a) before = np.identity(pow(2, i-1)) elem = np.kron(elem, before) elem = np.kron(elem, sigmacross + sigma) after = np.identity(pow(2, at_count-i)) elem = np.kron(elem, after) H_int += g * elem #------------------------------------------------ #------------------------------------------------------------------------------------------------------------------ return H_int
def TestSetCovariances(self,hyperparameters,X,Y): #compute test set covariances
[n, D] = X.shape;
ell = np.exp(hyperparameters[0:D]); # characteristic length scale
sf2 = np.exp(2*hyperparameters[D]); # signal variance
A = np.dot(sf2*(1+self.jitter),np.ones([np.size(Y,0),1]));
B = sf2*np.exp(-self.sq_dist2(np.dot(np.diagflat(1./ell),X.T),np.dot(np.diagflat(1./ell),Y.T))/2);#TODO diagflat
return [A,B] #TODO [A,B]
def test_project():
X = np.diagflat([-2, -1, 0, 1, 2])
# eigenvalues -2, -1, 0, 1, 2; eigenvectors are I
Xproj = ProjectPSD().fit_transform(X)
assert np.allclose(Xproj, np.diagflat([0, 0, 0, 1, 2]))
Xproj2 = ProjectPSD().fit(X).transform(X)
assert np.allclose(Xproj2, np.diagflat([0, 0, 0, 1, 2]))
Xproj3 = ProjectPSD(negatives_likely=True).fit(X).transform(X[:3, :])
assert np.allclose(Xproj3, np.zeros((3, 5)))
Xproj4 = ProjectPSD(negatives_likely=False).fit(X).transform(X[:3, :])
assert np.allclose(Xproj4, np.zeros((3, 5)))
Xproj5 = ProjectPSD(negatives_likely=True, copy=False, min_eig=.5) \
.fit_transform(X.copy())
assert np.allclose(Xproj5, np.diagflat([.5, .5, .5, 1, 2]))
Xproj6 = ProjectPSD(negatives_likely=True, copy=False, min_eig=.5) \
.fit(X.copy()).transform(X.copy())
assert np.allclose(Xproj6, np.diagflat([.5, .5, 0, 1, 2]))
assert_raises(TypeError, lambda: ProjectPSD().fit(X[:2, :]))
assert_raises(TypeError, lambda: ProjectPSD().fit_transform(X[:2, :]))
assert_raises(TypeError, lambda: ProjectPSD().fit(X).transform(X[:, :2]))
def eBBQ_participation2_H_params(s, cos_trunc = None, fock_trunc = None):
'''
returns the CHIs as MHz with anharmonicity alpha as the diagonal (with - sign)
f1: qubit dressed freq
f0: qubit linear freq (eigenmode)
and an overcomplete set of matrcieis
'''
import scipy; Planck = scipy.constants.Planck
f0s = s['freq'].values
Qs = s.loc[:,'modeQ']
LJs = s.loc[0,s.keys().str.contains('LJs')] # LJ in nH
EJs = (bbq.fluxQ**2/LJs/Planck*10**-9).astype(np.float) # EJs in GHz
PJ_Jsu = s.loc[:,s.keys().str.contains('pJ')] # EPR from Jsurf avg
PJ_Jsu_sum = PJ_Jsu.apply(sum, axis = 1) # sum of participations as calculated by avg surf current
PJ_glb_sum = (s['U_E'] - s['U_H'])/(2*s['U_E']) # sum of participations as calculated by global UH and UE
diff = (PJ_Jsu_sum-PJ_glb_sum)/PJ_glb_sum*100# debug
if 1: # Renormalize: to sum to PJ_glb_sum; so that PJs.apply(sum, axis = 1) - PJ_glb_sum =0
#TODO: figure out the systematic # print '% diff b/w Jsurf_avg & global Pj:'; display(diff)
PJs = PJ_Jsu.divide(PJ_Jsu_sum, axis=0).mul(PJ_glb_sum,axis=0)
else: PJs = PJ_Jsu
SIGN = s.loc[:,s.keys().str.contains('sign_')]
PJ = np.mat(PJs.values)
Om = np.mat(np.diagflat(f0s))
EJ = np.mat(np.diagflat(EJs.values))
CHI_O1= Om * PJ * EJ.I * PJ.T * Om * 1000 # MHz
CHI_O1= divide_diagonal_by_2(CHI_O1) # Make the diagonals alpha
f1s = f0s - np.diag(CHI_O1) # 1st order PT expect freq to be dressed down by alpha
if cos_trunc is not None:
f1s, CHI_ND, fzpfs, f0s = eBBQ_ND(f0s, PJ, Om, EJ, LJs, SIGN, cos_trunc = cos_trunc, fock_trunc = fock_trunc)
else: CHI_ND, fzpfs = None, None
return CHI_O1, CHI_ND, PJ, Om, EJ, diff, LJs, SIGN, f0s, f1s, fzpfs, Qs
def computePD(self, bodyUniqueId, jointIndices, desiredPositions, desiredVelocities, kps, kds, maxForces, timeStep):
numJoints = self._pb.getNumJoints(bodyUniqueId)
jointStates = self._pb.getJointStates(bodyUniqueId, jointIndices)
q1 = []
qdot1 = []
zeroAccelerations = []
for i in range (numJoints):
q1.append(jointStates[i][0])
qdot1.append(jointStates[i][1])
zeroAccelerations.append(0)
q = np.array(q1)
qdot=np.array(qdot1)
qdes = np.array(desiredPositions)
qdotdes = np.array(desiredVelocities)
qError = qdes - q
qdotError = qdotdes - qdot
Kp = np.diagflat(kps)
Kd = np.diagflat(kds)
p = Kp.dot(qError)
d = Kd.dot(qdotError)
forces = p + d
M1 = self._pb.calculateMassMatrix(bodyUniqueId,q1)
M2 = np.array(M1)
M = (M2 + Kd * timeStep)
c1 = self._pb.calculateInverseDynamics(bodyUniqueId, q1, qdot1, zeroAccelerations)
c = np.array(c1)
A = M
b = -c + p + d
qddot = np.linalg.solve(A, b)
tau = p + d - Kd.dot(qddot) * timeStep
maxF = np.array(maxForces)
forces = np.clip(tau, -maxF , maxF )
#print("c=",c)
return tau
def attention_prob_conv_matrix(Q, l):
assert l >= Q
m = np.diagflat([1.0] * l)
for i in range(1, Q):
m += np.diagflat([1.0] * (l - i), k=i)
m += np.diagflat([1.0] * (l - i), k=-i)
m = m / np.sum(m, axis=0)
return m
def build_attention_convolution_matrix(self, Q, l):
# straight up borrowed from Jencir's initial model
assert l >= Q
m = np.diagflat([1.] * l)
for i in range(1, Q):
m += np.diagflat([1.] * (l - i), k=i)
m += np.diagflat([1.] * (l - i), k=-i)
return m / np.sum(m, axis = 0)
def phi_2_sins(dim):
basic_sin = 1j*np.diagflat([-0.5]*(dim-1), 1) + 1j*np.diagflat([0.5]*(dim-1), -1)
basic_sin2 = np.diagflat([0.5]*dim, 0) + np.diagflat([-0.25]*(dim-2), 2) + np.diagflat([-0.25]*(dim-2), -2)
sin = np.tensordot(np.identity(dim), basic_sin, axes=0).swapaxes(1,2).reshape(dim*dim,-1)
sin2 = np.tensordot(np.identity(dim), basic_sin2, axes=0).swapaxes(1,2).reshape(dim*dim,-1)
return (sin, sin2)
def Smat(g):
"""
Return overlap matrix using triangular FEM basis.
"""
A=1.0/3.0*N.diagflat((g.Rx-g.Lx))
tmp=1.0/6.0*(g.Rx-g.x)
B=N.diagflat(tmp[0:g.NN-1],1)
return A+B+B.transpose()
def generate_laplacian_matrix(height, width):
N = height*width
a = np.diagflat(-4*np.ones(N), k=0)
b = np.diagflat(np.ones(N-1), k=1)
c = np.diagflat(np.ones(N-1), k=-1)
d = np.diagflat(np.ones(N-width), k=-width)
e = np.diagflat(np.ones(N-width), k=width)
return (a+b+c+d+e)*(height*width)
def __init__(self, skel, h):
self.h = h
self.skel = skel
ndofs = self.skel.ndofs
self.qhat = self.skel.q
self.Kp = np.diagflat([0.0] * 6 + [400.0] * (ndofs - 6))
self.Kd = np.diagflat([0.0] * 6 + [40.0] * (ndofs - 6))
self.preoffset = 0.0
def initialize_from_hdf5_file(file_name, structure, read_trajectory=True, initial_cut=1, final_cut=None, memmap=False):
import h5py
print("Reading data from hdf5 file: " + file_name)
trajectory = None
velocity = None
vc = None
reduced_q_vector = None
#Check file exists
if not os.path.isfile(file_name):
print(file_name + ' file does not exist!')
exit()
hdf5_file = h5py.File(file_name, "r")
if "trajectory" in hdf5_file and read_trajectory is True:
trajectory = hdf5_file['trajectory'][:]
if final_cut is not None:
trajectory = trajectory[initial_cut-1:final_cut]
else:
trajectory = trajectory[initial_cut-1:]
if "velocity" in hdf5_file:
velocity = hdf5_file['velocity'][:]
if final_cut is not None:
velocity = velocity[initial_cut-1:final_cut]
else:
velocity = velocity[initial_cut-1:]
if "vc" in hdf5_file:
vc = hdf5_file['vc'][:]
if final_cut is not None:
vc = vc[initial_cut-1:final_cut]
else:
vc = vc[initial_cut-1:]
if "reduced_q_vector" in hdf5_file:
reduced_q_vector = hdf5_file['reduced_q_vector'][:]
print("Load trajectory projected onto {0}".format(reduced_q_vector))
time = hdf5_file['time'][:]
supercell = hdf5_file['super_cell'][:]
hdf5_file.close()
if vc is None:
return dyn.Dynamics(structure=structure,
trajectory=trajectory,
velocity=velocity,
time=time,
supercell=np.dot(np.diagflat(supercell), structure.get_cell()),
memmap=memmap)
else:
return vc, reduced_q_vector, dyn.Dynamics(structure=structure,
time=time,
supercell=np.dot(np.diagflat(supercell), structure.get_cell()),
memmap=memmap)
def __init__(self, score_matrix, features):
self.source = score_matrix
u, e, vt = linalg.svds(self.source.raw_matrix.transpose().tocoo(), k = features)
self.normalization = self.source.normalization
e_inv = numpy.mat(numpy.diagflat(e)).getI()
e_sqrt = numpy.sqrt(numpy.diagflat(e))
u_mat = numpy.mat(u)
self.terms = u_mat * e_sqrt
self.documents = e_sqrt * e_inv * u_mat.transpose()
def build_Gamma(Gamma_diags):
"""Build Gamma (covariance) given the lists of diagonal values.
"""
Gamma = np.zeros((4, 4))
for k, diag in enumerate(Gamma_diags):
Gamma += np.diagflat(diag, k) # fill upper diagonals
if k > 0:
Gamma += np.diagflat(diag, -k) # fill lower diagonals
return Gamma
def linArd(hyp=None, x=None, z=None, hi=None, dg=None):
"""
Linear covariance function with Automatic Relevance Determination (ARD). The
covariance function is parameterized as:
k(x^p,x^q) = x^p'*inv(P)*x^q
where the P matrix is diagonal with ARD parameters ell_1^2,...,ell_D^2, where
D is the dimension of the input space. The hyperparameters are:
hyp = [ log(ell_1)
log(ell_2)
..
log(ell_D) ]
Note that there is no bias term; use covConst to add a bias.
"""
#report number of parameters
if x is None:
return 'D'
if z is None:
z = numpy.array([[]])
if dg is None:
dg = False
xeqz = numpy.size(z) == 0
ell = numpy.exp(hyp)
n, D = numpy.shape(x)
x = numpy.dot(x,numpy.diagflat(1./ell))
# compute inner products
if dg:
K = numpy.sum(x*x,1)
else:
if xeqz: # symmetric matrix Kxx
K = numpy.dot(x,x.T)
else: # cross covariances Kxz
z = numpy.dot(z,numpy.diagflat(1./ell))
K = numpy.dot(x,z.T)
if hi is not None: # derivatives
if hi > 0 and hi < D:
if dg:
K = -2*x[:,[i]]*x[:,[i]]
else:
if xeqz:
K = -2*numpy.dot(x[:,[i]],x[:,[i]].T)
else:
K = -2*numpy.dot(x[:,[i]],z[:,[i]].T)
else:
raise AttributeError('Unknown hyperparameter')
return K
def computePD(self, bodyUniqueId, jointIndices, desiredPositions, desiredVelocities, kps, kds,
maxForces, timeStep):
numBaseDofs = 0
numPosBaseDofs = 0
baseMass = self._pb.getDynamicsInfo(bodyUniqueId, -1)[0]
curPos, curOrn = self._pb.getBasePositionAndOrientation(bodyUniqueId)
q1 = []
qdot1 = []
zeroAccelerations = []
qError = []
if (baseMass > 0):
numBaseDofs = 6
numPosBaseDofs = 7
q1 = [curPos[0], curPos[1], curPos[2], curOrn[0], curOrn[1], curOrn[2], curOrn[3]]
qdot1 = [0] * numBaseDofs
zeroAccelerations = [0] * numBaseDofs
angDiff = [0, 0, 0]
qError = [
desiredPositions[0] - curPos[0], desiredPositions[1] - curPos[1],
desiredPositions[2] - curPos[2], angDiff[0], angDiff[1], angDiff[2]
]
numJoints = self._pb.getNumJoints(bodyUniqueId)
jointStates = self._pb.getJointStates(bodyUniqueId, jointIndices)
for i in range(numJoints):
q1.append(jointStates[i][0])
qdot1.append(jointStates[i][1])
zeroAccelerations.append(0)
q = np.array(q1)
qdot = np.array(qdot1)
qdes = np.array(desiredPositions)
qdotdes = np.array(desiredVelocities)
#qError = qdes - q
for j in range(numJoints):
qError.append(desiredPositions[j + numPosBaseDofs] - q1[j + numPosBaseDofs])
#print("qError=",qError)
qdotError = qdotdes - qdot
Kp = np.diagflat(kps)
Kd = np.diagflat(kds)
p = Kp.dot(qError)
d = Kd.dot(qdotError)
forces = p + d
M1 = self._pb.calculateMassMatrix(bodyUniqueId, q1)
M2 = np.array(M1)
M = (M2 + Kd * timeStep)
c1 = self._pb.calculateInverseDynamics(bodyUniqueId, q1, qdot1, zeroAccelerations)
c = np.array(c1)
A = M
b = -c + p + d
qddot = np.linalg.solve(A, b)
tau = p + d - Kd.dot(qddot) * timeStep
maxF = np.array(maxForces)
forces = np.clip(tau, -maxF, maxF)
#print("c=",c)
return tau
def KPMF(input_matrix, approx=50, iterations=30, learning_rate=.001, adjacency_width=5, adjacency_strength=.5):
A = input_matrix
Z = np.asarray(A > 0,dtype=np.int)
A1d = np.ravel(A)
mean = np.mean(A1d)
A = A-mean
K = approx
R = itr = iterations
l = learning_rate
N = A.shape[0]
M = A.shape[1]
U = np.random.randn(N,K)
V = np.random.randn(K,M)
#KPMF using gradient descent as per paper
#Kernelized Probabilistic Matrix Factorization: Exploiting Graphs and Side Information
#T. Zhou, H. Shan, A. Banerjee, G. Sapiro
#Using diffusion kernel
#U are the rows, we use an adjacency matrix CU to reprent connectivity
#This matrix connects rows +-adjacency_width
#V are the columns, connected columns are CV
#Operate on graph laplacian L, which is the degree matrix D - C
#Applying the diffusion kernel to L, this forms a spatial smoothness graph
bw = adjacency_width
#Use scipy.sparse.diags to generate band matrix with bandwidth = 2*adjacency_width+1
#Example of adjacency_width = 1, N = 4
#[1 1 0 0]
#[1 1 1 0]
#[0 1 1 1]
#[0 0 1 1]
print "Running KPMF with:"
print "learning rate=" + `l`
print "bandwidth=" + `bw`
print "beta=" + `b`
print "approximation rank=" + `K`
print "iterations=" + `R`
print ""
CU = sp.diags([1]*(2*bw+1),range(-bw,bw+1),shape=(N,N)).todense()
DU = np.diagflat(np.sum(CU,1))
CV = sp.diags([1]*(2*bw+1),range(-bw,bw+1),shape=(M,M)).todense()
DV = np.diagflat(np.sum(CV,1))
LU = DU - CU
LV = DV - CV
beta = adjacency_strength
KU = sl.expm(beta*LU)
KV = sl.expm(beta*LV)
SU = np.linalg.pinv(KU)
SV = np.linalg.pinv(KV)
for r in range(R):
for i in range(N):
for j in range(M):
if Z[i,j] > 0:
e = A[i,j] - np.dot(U[i,:],V[:,j])
U[i,:] = U[i,:] + l*(e*V[:,j] - np.dot(SU[i,:],U))
V[:,j] = V[:,j] + l*(e*U[i,:] - np.dot(V,SV[:,j]))
A_ = np.dot(U,V)
return A_+mean
def __init__( self, IC, fitness=5, dispersal=0, ndim=1, square=False,
verbose=False ):
"""
IC -- initial condition. A scale, list, or matrix, depending
on dimension of system (see ndim)
fitness -- The fitness on each patch. If scalar, each patch
has the same fitness. If numpy matrix of fitness values (same
dimension as ndim), different patches can have different
fitness values.
dispersal -- Dispersal parameter. As with fitness, can be either
scalar or numpy matrix.
ndim -- dimension of the system.
verbose -- output debugging and other chatter. (default=False)
"""
self.dispersal = dispersal # possibly a vector of matrix of different values
self.fitness = fitness # ditto
self.ndim = ndim
self.iter_number = 0
self._square = square
self._verbose = verbose
self._IC = IC
# change flag if creating dispersal matrix.
self._do_dispersal = False
if square:
self.sysdim = ndim * ndim
else:
self.sysdim = ndim
# now create patch population container
if not hasattr( IC, '__iter__' ):
# single initial population value will fill entire pop. matrix
if self._square:
self.population = np.empty( (ndim,ndim) )
self.population[:] = IC
# for single patch case
else:
self.population = IC
else:
# not much error catching here.
self.population = np.matrix( IC )
if self._square:
# create matrix of off-diagonals for dispersal computation
if not hasattr( self.dispersal, '__array__' ):
d = np.ones( self.ndim - 1 ) # clip one entry to account for
# off-diag position
self.D = np.diagflat( d, 1 ) + np.diagflat( d, -1 )
self.D = np.matrix( self.D ) # for matrix multiplication
self._do_dispersal = True
else:
print "Multiple values for dispersal not implemented yet!"
def make_quad_form(b1,b2,b3,b4,b5,b6,b7,b8,b9):
"""
Returns the quadratic form for the Fung model with the appropriate convention
for the material parameters.
The order of the strain components is assumed to be Theta, Z, R.
"""
normals = lin.symmetric(np.diagflat([b1,b2,b3]) +
2*np.diagflat([b4,b5],1) +
2*np.diagflat([b6],2))
return el.manual_stiffness(normals,np.diagflat([b7,b8,b9]))
def train(f_dataset_train, h_prev, m_prev, t_index):
inp_unit = f_dataset_train[:, 0].reshape(
input_size, 1) # input vector xt taken from image data
inp_whole = numpy.vstack(
(inp_unit, h_prev)) # input vector It combined from xt and ht-1
hidden_whole = numpy.dot(curr_model.get_wm_hidden(),
inp_whole) # raw hidden value for all gates zt
a_gate = numpy.tanh(hidden_whole[:curr_model.get_hidden_unit_size(), :]
) # split, squashed value from zt
i_gate = expit(hidden_whole[curr_model.get_hidden_unit_size():2 *
curr_model.get_hidden_unit_size(), :])
f_gate = expit(hidden_whole[2 * curr_model.get_hidden_unit_size():3 *
curr_model.get_hidden_unit_size(), :] + f_bias)
o_gate = expit(hidden_whole[3 * curr_model.get_hidden_unit_size():, :])
m_curr = i_gate * a_gate + f_gate * m_prev # memory cell value ct calculated from at, it, ft, and ct-1
h_curr = o_gate * numpy.tanh(
m_curr) # current timestep output value calculated from ot and ct
if f_dataset_train[:,
1:].size == 0: # on the last step, calculate score, errhc
score = numpy.dot(curr_model.get_wm_score(),
h_curr) # output layer producing 24 length vector s
score_prob = numpy.exp(score) / numpy.sum(
numpy.exp(score)) # softmax layer producing probability vector P
loss = -numpy.log(
score_prob[t_index]) # negative log loss error value E
jac = numpy.diagflat(score_prob) - numpy.dot(
score_prob, score_prob.T) # matrix J for deriving softmax layer
err_score = numpy.dot(
jac, score_prob -
target_array[t_index]) # error vector ds from J and audio target
err_wm_score = numpy.dot(
err_score, h_curr.T) # matrix error dws calculated from ds
curr_model.set_wm_score(
curr_model.get_wm_score() -
curr_model.get_learning_rate() * err_wm_score) # ws update
err_h_curr = numpy.dot(
curr_model.get_wm_score().T,
err_score) # error vector dht for deriving previous steps
err_m_curr = 0 # last step has no next memory cell, dct = 0
err_wm_hidden = 0 # last step has 0 accumulated weight error, dw = 0
else: # if not on last step, send ht, ct for forward prop, wait for dht, dct, dw for backward prop
err_h_curr, err_m_curr, err_wm_hidden, loss = train(
f_dataset_train[:, 1:], h_curr, m_curr, t_index)
err_m_curr = err_m_curr + (err_h_curr * o_gate *
(1 - numpy.power(numpy.tanh(m_curr), 2))
) # accumulative dct
err_a_gate = err_m_curr * i_gate # error vector dat calculated from dct
err_i_gate = err_m_curr * a_gate # error vector dat calculated from dct
err_f_gate = err_m_curr * m_prev # error vector dat calculated from dct
err_o_gate = err_h_curr * numpy.tanh(
m_curr) # error vector dot calculated from dht
err_m_prev = err_m_curr * f_gate # error vector dct-1 calculated from dct
# calculate error vector for raw, unsplit, unsquashed hidden values for all at, it, ft, and ot
err_a_inp = err_a_gate * (1 - numpy.power(
numpy.tanh(hidden_whole[:curr_model.get_hidden_unit_size(), :]), 2))
err_i_inp = err_i_gate * i_gate * (1 - i_gate)
err_f_inp = err_f_gate * f_gate * (1 - f_gate)
err_o_inp = err_o_gate * o_gate * (1 - o_gate)
err_inp_whole = numpy.vstack(
(err_a_inp, err_i_inp, err_f_inp,
err_o_inp)) # dzt concated from hidden value errors
err_wm_hidden = err_wm_hidden + numpy.dot(
err_inp_whole, inp_whole.T) # accumulative dw for weight update
err_inp_unit = numpy.dot(curr_model.get_wm_hidden().T,
err_inp_whole) # whole input error vector dIt
err_h_prev = err_inp_unit[input_size:, 0].reshape(
(curr_model.get_hidden_unit_size(), 1)) # dht-1 split from dIt
return err_h_prev, err_m_prev, err_wm_hidden, loss # returns dht-1, dct-1, dw for backprop
def compute_differentials(grey_level_matrix, diagonal_neighbors=True):
"""Computes differences in greylevels for neighboring grid points.
First part of 'step 4' in the paper.
Returns n x n x 8 rank 3 array for an n x n grid (if diagonal_neighbors == True)
The n x nth coordinate corresponds to a grid point. The eight values are
the differences between neighboring grid points, in this order:
upper left
upper
upper right
left
right
lower left
lower
lower right
Args:
grey_level_matrix (numpy.ndarray): grid of values sampled from image
diagonal_neighbors (Optional[boolean]): whether or not to use diagonal
neighbors (default True)
Returns:
a n x n x 8 rank 3 numpy array for an n x n grid (if diagonal_neighbors == True)
Examples:
>>> img = gis.preprocess_image('https://pixabay.com/static/uploads/photo/2012/11/28/08/56/mona-lisa-67506_960_720.jpg')
>>> window = gis.crop_image(img)
>>> grid = gis.compute_grid_points(img, window=window)
>>> grey_levels = gis.compute_mean_level(img, grid[0], grid[1])
>>> gis.compute_differentials(grey_levels)
array([[[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 1.82683143e-03, -0.00000000e+00,
2.74085276e-01, 1.24737821e-01],
[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
-1.82683143e-03, 2.15563930e-03, 2.72258444e-01,
1.22910990e-01, 3.48522956e-01],
[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
-2.15563930e-03, 1.16963917e-01, 1.20755351e-01,
3.46367317e-01, 1.99638513e-01],
[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
-1.16963917e-01, 1.32128118e-01, 2.29403399e-01,
8.26745956e-02, 1.16455050e-01],
...
"""
right_neighbors = -np.concatenate(
(np.diff(grey_level_matrix), np.zeros(
grey_level_matrix.shape[0]).reshape(
(grey_level_matrix.shape[0], 1))),
axis=1)
left_neighbors = -np.concatenate(
(right_neighbors[:, -1:], right_neighbors[:, :-1]), axis=1)
down_neighbors = -np.concatenate(
(np.diff(grey_level_matrix,
axis=0), np.zeros(grey_level_matrix.shape[1]).reshape(
(1, grey_level_matrix.shape[1]))))
up_neighbors = -np.concatenate(
(down_neighbors[-1:], down_neighbors[:-1]))
if diagonal_neighbors:
# this implementation will only work for a square (m x m) grid
diagonals = np.arange(-grey_level_matrix.shape[0] + 1,
grey_level_matrix.shape[0])
upper_left_neighbors = sum([
np.diagflat(
np.insert(np.diff(np.diag(grey_level_matrix, i)), 0, 0), i)
for i in diagonals
])
lower_right_neighbors = -np.pad(upper_left_neighbors[1:, 1:],
(0, 1),
mode='constant')
# flip for anti-diagonal differences
flipped = np.fliplr(grey_level_matrix)
upper_right_neighbors = sum([
np.diagflat(np.insert(np.diff(np.diag(flipped, i)), 0, 0), i)
for i in diagonals
])
lower_left_neighbors = -np.pad(upper_right_neighbors[1:, 1:],
(0, 1),
mode='constant')
return np.dstack(
np.array([
upper_left_neighbors, up_neighbors,
np.fliplr(upper_right_neighbors), left_neighbors,
right_neighbors,
np.fliplr(lower_left_neighbors), down_neighbors,
lower_right_neighbors
]))
return np.dstack(
np.array([
up_neighbors, left_neighbors, right_neighbors, down_neighbors
]))
def second_order_solver(FF, GG, HH):
from scipy.linalg import qz
from dolo.numeric.extern.qz import qzdiv
from numpy import array, mat, c_, r_, eye, zeros, real_if_close, diag, allclose, where, diagflat
from numpy.linalg import solve
Psi_mat = array(FF)
Gamma_mat = array(-GG)
Theta_mat = array(-HH)
m_states = FF.shape[0]
Xi_mat = r_[c_[Gamma_mat, Theta_mat], c_[eye(m_states),
zeros((m_states, m_states))]]
Delta_mat = r_[c_[Psi_mat, zeros((m_states, m_states))], c_[zeros(
(m_states, m_states)), eye(m_states)]]
AAA, BBB, Q, Z = qz(Delta_mat, Xi_mat)
Delta_up, Xi_up, UUU, VVV = [real_if_close(mm) for mm in (AAA, BBB, Q, Z)]
Xi_eigval = diag(Xi_up) / where(diag(Delta_up) > TOL, diag(Delta_up), TOL)
Xi_sortindex = abs(Xi_eigval).argsort()
# (Xi_sortabs doesn't really seem to be needed)
Xi_sortval = Xi_eigval[Xi_sortindex]
Xi_select = slice(0, m_states)
stake = (abs(Xi_sortval[Xi_select])).max() + TOL
Delta_up, Xi_up, UUU, VVV = qzdiv(stake, Delta_up, Xi_up, UUU, VVV)
try:
# check that all unused roots are unstable
assert abs(Xi_sortval[m_states]) > (1 - TOL)
# check that all used roots are stable
assert abs(Xi_sortval[Xi_select]).max() < 1 + TOL
except:
raise BKError('generic')
# check for unit roots anywhere
# assert (abs((abs(Xi_sortval) - 1)) > TOL).all()
Lambda_mat = diagflat(Xi_sortval[Xi_select])
VVVH = VVV.T
VVV_2_1 = VVVH[m_states:2 * m_states, :m_states]
VVV_2_2 = VVVH[m_states:2 * m_states, m_states:2 * m_states]
UUU_2_1 = UUU[m_states:2 * m_states, :m_states]
PP = -solve(VVV_2_1, VVV_2_2)
# slightly different check than in the original toolkit:
assert allclose(real_if_close(PP), PP.real)
PP = PP.real
## end of solve_qz!
return [Xi_sortval[Xi_select], PP]
detunings = np.zeros(states_count)
for i in range(2 * n - 1):
if i >= states_count:
break
if i < n:
detunings[i] = (i - s) * rf_freq - (eigenvalues[i] -
eigenvalues[s])
else:
detunings[i] = rf_freq - (eigenvalues[i - n + 1] -
eigenvalues[i]) + detunings[i - n + 1]
detunings *= -1
energies.append(detunings)
# Construct a new Hamiltonian using the no-RF detunings along the diagonal.
hamiltonian_with_rf = e_rf * mat_2_combination + np.diagflat(detunings)
eigenvalues, eigenvectors = np.linalg.eigh(hamiltonian_with_rf)
energies_with_rf.append(eigenvalues)
energies = np.array(energies)
energies_with_rf = np.array(energies_with_rf)
fig, (ax1, ax2) = plt.subplots(
1,
2,
sharex='all',
sharey='all',
figsize=(8, 5),
)
def decomposition(A):
L = np.tril(A, -1)
D = np.diagflat(np.diag(A))
R = np.triu(A, 1)
return (L, D, R)
# Realized marginals mapping into standard Student t realizations
u_stocks = zeros((i_, t_))
epsi_tilde_stocks = zeros((i_, t_))
for i in range(i_):
# u_stocks([i,:])=min((t.cdf((epsi_stocks[i,:]-mu_marg[i])/sqrt(sig2_marg[i]),nu_marg[i]),0.999))
u_stocks[i, :] = tstu.cdf(
(epsi_stocks[i, :] - mu_marg[i]) / sqrt(sig2_marg[i]), nu_marg[i])
epsi_tilde_stocks[i, :] = tstu.ppf(u_stocks[i, :],
nu_copula) # Student t realizations
# Correlation matrix characterizing the t copula estimation
# approximate the fit to normal in case of badly scaled warnings
_, sig2, _ = MaxLikelihoodFPLocDispT(epsi_tilde_stocks, p, 1e9, 1e-6, 1)
rho2 = np.diagflat(diag(sig2)**(-1 / 2)) @ sig2 @ np.diagflat(
diag(sig2)**(-1 / 2))
# Shrink the correlation matrix towards a low-rank-diagonal structure
rho2, beta, *_ = FactorAnalysis(rho2, array([[0]]), k_)
rho2, beta = np.real(rho2), np.real(beta)
# Monte Carlo scenarios for each path node from the t copula
Epsi_tilde_hor = zeros((i_, m_, j_))
optionT = namedtuple('option', 'dim_red stoc_rep')
optionT.dim_red = 0
optionT.stoc_rep = 0
for m in range(m_):
Epsi_tilde_hor[:, m, :] = Tscenarios(nu_copula, zeros(
(i_,
1)), rho2, j_, optionT) # We simulate scenarios one node at a time
def hessian(self, w):
self._num_calls += 1
prob = self.y_pred(self._X @ w)
return 1 / self.N * self._X.T @ np.diagflat(
np.multiply(prob, (1 - prob))) @ self._X
def backward(self, dinputs: np.ndarray):
self.dinputs = np.empty_like(dinputs)
for i, (output, dinput) in enumerate(zip(self.outputs, dinputs)):
output = output.reshape(-1, 1)
jacobian_matrix = np.diagflat(output) - np.dot(output, output.T)
self.dinputs[i] = np.dot(jacobian_matrix, dinput)
def time_diagflat_l50_l50(self):
np.diagflat([self.l50, self.l50])
Vsector = data_sectors[:k_, :] # values Z = (Vsector[:, 1:] - Vsector[:, :-1]) / Vsector[:, :-1] # ## Compute statistics of the joint distribution of X,Z [m_XZ, s2_XZ] = FPmeancov(r_[X, Z], p) s2_X = s2_XZ[:n_, :n_] s_XZ = s2_XZ[:n_, n_:n_ + k_] s2_Z = s2_XZ[n_:n_ + k_, n_:n_ + k_] # ## Solve generalized regression LFM # ## set inputs for quadratic programming problem # + d = np.diagflat(1 / diag(s2_X)) pos = d @ s_XZ g = -pos.flatten() q = kron(s2_Z, d) q_, _ = q.shape # set constraints a_eq = ones((1, n_ * k_)) / (n_ * k_) b_eq = array([[1]]) lb = 0.8 * ones((n_ * k_, 1)) ub = 1.2 * ones((n_ * k_, 1)) # compute optimal loadings b = quadprog(q, g, a_eq, b_eq, lb, ub) b = np.array(b)
def softmax_d(softmax):
s = softmax.reshape(-1,1)
return np.diagflat(s) - np.dot(s, s.T)
def garch1f4(x, eps, df):
## Fit a GARCH(1,1) model with student-t errors
# INPUTS
# x : [vector] (T x 1) data generated by a GARCH(1,1) process
# OPS
# q : [vector] (4 x 1) parameters of the GARCH(1,1) process
# qerr : [vector] (4 x 1) standard error of parameter estimates
# hf : [scalar] current conditional heteroskedasticity estimate
# hferr : [scalar] standard error on hf
# NOTE
# o Uses a conditional t-distribution with fixed degrees of freedom
# o Originally written by Olivier Ledoit, 4/28/1997
# o Difference with garch1f: errors come from the score alone
# Parameters
gold = (1 + sqrt(5)) / 2 # step size increment
tol1 = 1e-7 # for termination criterion
tol2 = 1e-7 # for closeness to boundary
big = 2 # for making the hessian negative definite
maxiter = 50 # maximum number of iterations
n = 30 # number of points on the grid
# Rescale
y = (x.flatten() - mean(x.flatten()))**2
t = len(y)
scale = sqrt(mean(y**2))
y = y / scale
s = mean(y)
# Grid search
[ag, bg] = meshgrid(linspace(0, 1 - eps, n), linspace(0, 1 - eps, n))
cg = np.maximum(s * (1 - ag - bg), 0)
likeg = -np.Inf * ones((n, n))
for i in range(n):
for j in range(n - i):
h = filter(array([0, ag[i, j]]), array([1, -bg[i, j]]), y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter(array([0, cg[i, j]]), array([1, -bg[i, j]]), ones(t))
likeg[i, j] = -npsum(log(h) + (df + 1) * log(1 + y / h / df))
maxlikeg = npmax(likeg)
maxima = where(likeg == maxlikeg) ##ok<MXFND>
# Initialize optimization
a = r_[cg[maxima], ag[maxima], bg[maxima]]
best = 0
da = 0
# term = 1
# negdef = 0
iter = 0
# Begin optimization loop
while iter < maxiter:
iter = iter + 1
# New parameter1
a = a + gold**best * da
# Conditional variance
h = filter([0, a[1]], [1, -a[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, a[0]], [1, -a[2]], ones(t))
# Likelihood
if (any(a < 0) or ((a[1] + a[2]) > 1 - eps)):
like = -np.Inf
else:
like = -npsum(log(h) + (df + 1) * log(1 + y / h / df))
# Gradient
GG = r_['-1',
filter([0, 1], [1, -a[2]], ones(t))[..., newaxis],
filter([0, 1], [1, -a[2]],
y * (df - 2) / df)[..., newaxis],
filter([0, 1], [1, -a[2]], h)[..., newaxis]]
g1 = ((df + 1) * (y / (y + df * h)) - 1) / h
G = GG * repeat(g1.reshape(-1, 1), 3, axis=1)
gra = npsum(G, axis=0)
# Hessian
GG2 = GG[:,
[0, 1, 2, 0, 1, 2, 0, 1, 2]] * GG[:,
[0, 0, 0, 1, 1, 1, 2, 2, 2]]
g2 = -((df + 1) *
(y /
(y + df * h)) - 1) / h**2 - (df *
(df + 1)) * (y /
(y + df * h)**2 / h)
HH = zeros((t, 9))
HH[:, 2] = filter([0, 1], [1, -a[2]], GG[:, 0])
HH[:, 6] = HH[:, 2]
HH[:, 5] = filter([0, 1], [1, -a[2]], GG[:, 1])
HH[:, 7] = HH[:, 5]
HH[:, 8] = filter([0, 2], [1, -a[2]], GG[:, 2])
H = GG2 * repeat(g2.reshape(-1, 1), 9, axis=1) + HH * repeat(
g1.reshape(-1, 1), 9, axis=1)
hes = reshape(npsum(H, axis=0), (3, 3), 'F')
# Negative definite
d, u = eig(hes)
# d = diagflat(d)
if any(d > 0):
negdef = 0
d = min(d, max(d[d < 0]) / big)
hes = u @ diagflat(d) @ u.T
else:
negdef = 1
# Direction
da = -gra.dot(pinv(hes))
# Termination criterion
term = da @ gra.T
if (term < tol1) and negdef:
break
# Step search
best = 0
newa = a + gold**(best - 1) * da
if (any(newa < 0) or (newa[1] + newa[2] > 1 - eps)):
left = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
left = -sum(log(h) + (df + 1) * log(1 + y / h / df))
newa = a + gold**best * da
if (any(newa < 0) or (newa[1] + newa[2] > 1 - eps)):
center = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
center = -sum(log(h) + (df + 1) * log(1 + y / h / df))
newa = a + gold**(best + 1) * da
if (any(newa < 0) or (newa[1] + newa[2] > 1 - eps)):
right = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
right = -sum(log(h) + (df + 1) * log(1 + y / h / df))
if all(like > array([left, center, right])) or all(
left > array([center, right])):
while True:
best = best - 1
center = left
newa = a + gold**(best - 1) * da
if (any(newa < 0) or (newa[1] + newa[2] > 1 - eps)):
left = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
left = -sum(log(h) + (df + 1) * log(1 + y / h / df))
if all(center >= array([like, left])):
break
elif all(right > array([left, center])):
while True:
best = best + 1
center = right
newa = a + gold**(best + 1) * da
if (any(newa < 0) or (newa[1] + newa[2]) > 1 - eps):
right = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
right = -npsum(log(h) + (df + 1) * log(1 + y / h / df))
if center > right:
break
# If stuck at boundary then stop
if (center == like) and (any(a < tol2) or (a[1] + a[2]) > 1 - tol2):
break
# End of optimization loop
a[a < tol2] = zeros(len(a[a < tol2]))
if a[1] + a[2] > 1 - tol2:
if a[1] < 1 - tol2:
a[1] = a[1] + (1 - a[1] - a[2])
else:
a[2] = a[2] + (1 - a[1] - a[2])
# Estimation error and volatility forecast
# aerr=inv(G.T@G)
tmp = (G.T @ G)
aerr = tmp.dot(pinv(eye(tmp.shape[0])))
hf = a[0] + a[1] * y[t - 1] * (df - 2) / df + a[2] * h[t - 1]
gf = r_[1, y[t - 1], h[t - 1]] + a[2] * GG[t - 1, :]
hferr = gf @ aerr @ gf.T
aerr = diagflat(aerr).T
# Revert to original scale
a[0] = a[0] * scale
aerr[0] = aerr[0] * scale**2
hf = hf * scale
hferr = hferr * scale**2
aerr = sqrt(aerr)
hferr = sqrt(hferr)
q = a
qerr = aerr
return q, qerr, hf, hferr
def __init__(self,
grid=None,
edges=None,
origin=None,
delta=None,
metadata=None,
**kwargs):
"""
Create a Grid object from data.
From a numpy.histogramdd()::
grid,edges = numpy.histogramdd(...)
g = Grid(grid,edges=edges)
From an arbitrary grid::
g = Grid(grid,origin=origin,delta=delta)
From a saved file::
g = Grid(filename)
or
g = Grid()
g.load(filename)
:Arguments:
grid
histogram or density, defined on numpy nD array
edges
list of arrays, the lower and upper bin edges along the axes
(both are output by numpy.histogramdd())
origin
cartesian coordinates of the center of grid[0,0,...,0]
delta
Either n x n array containing the cell lengths in each dimension,
or n x 1 array for rectangular arrays.
metadata
a user defined dictionary of arbitrary values
associated with the density; the class does not touch
metadata[] but stores it with save()
interpolation_spline_order
order of interpolation function for resampling; cubic splines = 3 [3]
"""
# file formats are guess from extension == lower case key
self._exporters = {
'DX': self._export_dx,
'PICKLE': self._export_python,
'PYTHON': self._export_python, # compatibility
}
self._loaders = {
'DX': self._load_dx,
'PLT': self._load_plt,
'PICKLE': self._load_python,
'PYTHON': self._load_python, # compatibility
}
if metadata is None:
metadata = {}
self.metadata = metadata # use this to record arbitrary data
self.__interpolated = None # cache for interpolated grid
self.__interpolation_spline_order = kwargs.pop(
'interpolation_spline_order', 3)
self.interpolation_cval = None # default to using min(grid)
if type(grid) is str:
# read from a file
self.load(grid)
elif not (grid is None or edges is None):
# set up from histogramdd-type data
self.grid = numpy.asarray(grid)
self.edges = edges
self._update()
elif not (grid is None or origin is None or delta is None):
# setup from generic data
origin = numpy.squeeze(origin)
delta = numpy.squeeze(delta)
N = grid.ndim
assert (N == len(origin))
if delta.shape == (N, N):
if numpy.any(delta - numpy.diag(delta)):
raise NotImplementedError(
"Non-rectangular grids are not supported.")
elif delta.shape == (N, ):
delta = numpy.diagflat(delta)
elif delta.shape == ():
delta = numpy.diagflat(N * [delta])
else:
raise ValueError('delta = %r has the wrong shape' % delta)
# note that origin is CENTER so edges must be shifted by -0.5*delta
self.edges = [
origin[dim] + (numpy.arange(m + 1) - 0.5) * delta[dim, dim]
for dim, m in enumerate(grid.shape)
]
self.grid = numpy.asarray(grid)
self._update()
else:
# empty, must manually populate with load()
#print "Setting up empty grid object. Use Grid.load(filename)."
pass
def garch2f8(y, c1, a1, b1, y1, h1, c2, a2, b2, y2, h2, df):
## Off-diagonal parameter estimation in bivariate GARCH(1,1) when diagonal parameters are given.
# INPUTS
# y : [vector] (T x 1) data generated by a GARCH(1,1) process
# OPS
# q : [vector] (4 x 1) parameters of the GARCH(1,1) process
# qerr : [vector] (4 x 1) standard error of parameter estimates
# hf : [scalar] current conditional heteroskedasticity estimate
# hferr : [scalar] standard error on hf
# NOTE
# o Originally written by Olivier Ledoit, 4/28/1997
# o Uses a conditional t-distribution with fixed degrees of freedom
# o Steepest Ascent on boundary, Hessian off boundary, no grid search
# Parameters
gold = (1 + sqrt(5)) / 2 # step size increment
tol1 = 1e-7 # for termination criterion
tol2 = 1e-7 # for closeness to boundary
big = 2 # for making the hessian negative definite
maxiter = 50 # maximum number of iterations
# n=30 # number of points on the grid
# Prepare
t = len(y)
y1 = y1.flatten()
y2 = y2.flatten()
y = y.flatten()
s = mean(y)
# s1=mean((y1))
# s2=mean((y2))
h1 = h1.flatten()
h2 = h2.flatten()
# Bounds
low = r_[-sqrt(c1 * c2), 0, 0] + tol2
high = r_[sqrt(c1 * c2), sqrt(a1 * a2), sqrt(b1 * b2)] - tol2
# Starting Point
a0 = 0.9 * sqrt(a1 * a2)
b0 = 0.9 * sqrt(b1 * b2)
c0 = mean(y) * (1 - a0 - b0) * (df - 2) / df
c0 = sign(c0) * min(abs(c0), 0.9 * sqrt(c1 * c2))
# Initialize optimization
a = r_[c0, a0, b0]
best = 0
da = 0
# term=1
# negdef=0
iter = 0
# Begin optimization loop
while iter < maxiter:
iter = iter + 1
# New parameter
# olda = a
a = a + gold**best * da
# Conditional variance
h = filter([0, a[1]], [1, -a[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, a[0]], [1, -a[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
# Likelihood
if (any(a < low) or any(a > high)):
like = -np.Inf
else:
# like=-sum(log(h)+y/h))
# like=-sum(log(h)+(df+1)*log(1+y/h/df))
if any(d <= 0) or any(1 + z / d / df <= 0):
like = -np.Inf
else:
like = -sum(log(d) + (2 + df) * log(1 + z / d / df)) / 2
# Gradient
GG = r_['-1',
filter([0, 1], [1, -a[2]], ones(t))[..., newaxis],
filter([0, 1], [1, -a[2]],
y * (df - 2) / df)[..., newaxis],
filter([0, 1], [1, -a[2]], h)[..., newaxis]]
g1 = h / d + (2 + df) * y / (z + d * df) - (2 + df) * h * z / (
z + d * df) / d
G = GG * repeat(g1.reshape(-1, 1), 3, axis=1)
gra = npsum(G, axis=0)
# Hessian
GG2 = GG[:,
[0, 1, 2, 0, 1, 2, 0, 1, 2]] * GG[:,
[0, 0, 0, 1, 1, 1, 2, 2, 2]]
g2 = 1 / d + 2 * h ** 2 / d ** 2 - (2 + df) * y / (z + d * df) ** 2 * (-2 * y - 2 * df * h) \
- (2 + df) * z / (z + d * df) / d + 2 * (2 + df) * h * y / (z + d * df) / d \
+ (2 + df) * h * z / (z + d * df) ** 2 / d * (-2 * y - 2 * df * h) \
- 2 * (2 + df) * h ** 2 * z / (z + d * df) / d ** 2
HH = zeros((t, 9))
HH[:, 2] = filter([0, 1], [1, -a[2]], GG[:, 0])
HH[:, 6] = HH[:, 2]
HH[:, 5] = filter([0, 1], [1, -a[2]], GG[:, 1])
HH[:, 7] = HH[:, 5]
HH[:, 8] = filter([0, 2], [1, -a[2]], GG[:, 2])
H = GG2 * repeat(g2.reshape(-1, 1), 9, axis=1) + HH * repeat(
g1.reshape(-1, 1), 9, axis=1)
hes = reshape(npsum(H, axis=0), (3, 3), 'F')
# Negative definite
val, u = eig(hes)
if all(val > 0):
hes = -eye(3)
negdef = 0
elif any(val > 0):
negdef = 0
val = minimum(val, max(val[val < 0]) / big)
hes = u @ diagflat(val) @ u.T
else:
negdef = 1
# Steepest Ascent or Newton
if any(a == low) or any(a == high):
da = -((gra @ gra.T) / (gra @ hes @ gra.T)) * gra
else:
da = -gra.dot(pinv(hes))
# Termination criterion
term = da @ gra.T
if ((term < tol1) and negdef):
break
# If you are on the boundary and want to get out, slide along
da[(a == low) & (da < 0)] = zeros(da[(a == low) & (da < 0)].shape)
da[(a == high) & (da > 0)] = zeros(da[(a == high) & (da > 0)].shape)
# If you are stuck in a corner, terminate too
if all(da == 0):
break
# Go no further than next boundary
hit = r_[(low[da != 0] - a[da != 0]) / da[da != 0],
(high[da != 0] - a[da != 0]) / da[da != 0]]
hit = hit[hit > 0]
da = min(r_[hit, 1]) * da
# Step search
best = 0
newa = a + gold**(best - 1) * da
if (any(newa < low) or any(newa > high)):
left = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
if any(d <= 0) or any(1 + z / d / df <= 0):
left = -np.Inf
else:
left = -sum(log(d) + (2 + df) * log(1 + z / d / df)) / 2
newa = a + gold**best * da
if (any(newa < low) or any(newa > high)):
center = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
if any(d <= 0) or any(1 + z / d / df <= 0):
center = -np.Inf
else:
center = -sum(log(d) + (2 + df) * log(1 + z / d / df)) / 2
newa = a + gold**(best + 1) * da
if (any(newa < low) or any(newa > high)):
right = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
if any(d <= 0) or any(1 + z / d / df <= 0):
right = -np.Inf
else:
right = -sum(log(d) + (2 + df) * log(1 + z / d / df)) / 2
if all(like > array([left, center, right])) or all(
left > array([center, right])):
while True:
best = best - 1
center = left
newa = a + gold**(best - 1) * da
if (any(newa < low) or any(newa > high)):
left = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
if any(d <= 0) or any(1 + z / d / df <= 0):
left = -np.Inf
else:
left = -sum(log(d) +
(2 + df) * log(1 + z / d / df)) / 2
if all(center >= [like, left]):
break
elif all(right > array([left, center])):
while True:
best = best + 1
center = right
newa = a + gold**(best + 1) * da
if (any(newa < low) or any(newa > high)):
right = -np.Inf
else:
h = filter([0, newa[1]], [1, -newa[2]], y * (df - 2) / df, zi=array([s * (df - 2) / df]))[0] \
+ filter([0, newa[0]], [1, -newa[2]], ones(t))
d = h1 * h2 - h**2
z = h2 * y1 + h1 * y2 - 2 * h * y
if any(d <= 0) or any(1 + z / d / df <= 0):
right = -np.Inf
else:
right = -npsum(
log(d) + (2 + df) * log(1 + z / d / df)) / 2
if center > right:
break
q = a
return q
plt.text(estimate_xs[-1],
estimate_ys[-1],
str(i + 5),
fontsize=10)
plt.scatter(estimate_xs,
estimate_ys,
s=50,
c='k',
marker='.',
label='Landmark Estimate')
plt.title('Graph SLAM with known correspondences')
plt.legend()
plt.xlim((-2.0, 5.5))
plt.ylim((-7.0, 7.0))
if __name__ == "__main__":
# Dataset 1
dataset = "../0.Dataset1"
start_frame = 800
end_frame = 2000
N_iterations = 1
# State covariance matrix
R = np.diagflat(np.array([5, 5, 20]))**2
# Measurement covariance matrix
Q = np.diagflat(np.array([100.0, 100.0, 1e16]))**2
graph_slam = GraphSLAM(dataset, start_frame, end_frame, N_iterations, R, Q)
plt.show()
def softmax_derivative(self, x):
# Reshape the 1-d softmax to 2-d so that np.dot will do the matrix multiplication
z = x.reshape(-1, 1)
return np.diagflat(z) - np.dot(z, z.T)
lam = (log(2)) / 120 # half life 4 months
flex_prob = exp(-lam * arange(t_, 1 + -1, -1)).reshape(1, -1)
flex_prob = flex_prob / npsum(flex_prob)
typ = namedtuple('typ', 'Entropy')
typ.Entropy = 'Exp'
ens = EffectiveScenarios(flex_prob, typ)
# -
# ## Twist fix for non-synchroneity in HFP
# +
print('Performing the twist fix for non-synchroneity')
# (step 1-2) HFP MEAN/COVARIANCE/CORRELATION
HFPmu, HFPcov = FPmeancov(epsi, flex_prob)
HFPc2 = np.diagflat(diag(HFPcov)**(-1 / 2)) @ HFPcov @ np.diagflat(
diag(HFPcov)**(-1 / 2))
# (step 3) TARGET CORRELATIONS
l = 10 # number of lags
flex_prob_l = flex_prob[[0], l:]
flex_prob_l = flex_prob_l / npsum(flex_prob_l)
# concatenate the daily log-returns
y1, y2 = zeros(t_), zeros(t_)
for t in range(l, t_):
y1[t] = sum(ret1[0, t - l:t])
y2[t] = sum(ret2[0, t - l:t])
y1 = y1[l:]
def subsolve(self, m, n, epsimin, alfa, beta, p0, q0, P, Q):
# init variables
een = np.ones((n, 1))
eem = np.ones((m, 1))
epsi = 1
epsvecn = epsi * np.ones((n, 1))
epsvecm = epsi * np.ones((m, 1))
x = 0.5 * (alfa + beta)
y = np.ones((m, 1))
z = np.ones((1, 1))
lam = np.ones((m, 1))
xsi = np.reciprocal(x - alfa)
xsi = np.maximum(xsi, een)
eta = np.reciprocal(beta - x)
eta = np.maximum(eta, een)
mu = np.maximum(eem, 0.5 * self.c)
zet = np.ones((1, 1))
s = np.ones((m, 1))
itera = 0
while epsi > epsimin:
epsvecn.fill(epsi)
epsvecm.fill(epsi)
ux1 = self.upp - x
xl1 = x - self.low
ux2 = ux1 * ux1
xl2 = xl1 * xl1
uxinv1 = np.reciprocal(ux1)
xlinv1 = np.reciprocal(xl1)
plam = p0 + np.transpose(P) * lam
qlam = q0 + np.transpose(Q) * lam
gvec = np.dot(P, uxinv1) + np.dot(Q, xlinv1)
dpsidx = plam / ux2 - qlam / xl2
rex = dpsidx - xsi + eta
rey = self.c + self.d * y - mu - lam
rez = self.a0 - zet - np.transpose(self.a) * lam
relam = gvec - np.dot(self.a, z) - y + s - self.b
rexsi = xsi * (x - alfa) - epsvecn
reeta = eta * (beta - x) - epsvecn
remu = mu * y - epsvecm
rezet = zet * z - epsi
res = lam * s - epsvecm
residu1 = np.concatenate((rex, rey, rez))
residu2 = np.concatenate((relam, rexsi, reeta, remu, rezet, res))
residu = np.concatenate((residu1, residu2))
#compute norm & norm
residunorm = np.sqrt((residu * residu).sum())
residumax = np.max(np.absolute(residu))
ittt = 0
while residumax > 0.9 * epsi and ittt < 200:
ittt = ittt + 1
itera = itera + 1
ux1 = self.upp - x
xl1 = x - self.low
ux2 = ux1 * ux1
xl2 = xl1 * xl1
ux3 = ux1 * ux2
xl3 = xl1 * xl2
uxinv1 = np.reciprocal(ux1)
xlinv1 = np.reciprocal(xl1)
uxinv2 = np.reciprocal(ux2)
xlinv2 = np.reciprocal(xl2)
plam = p0 + np.dot(np.transpose(P), lam)
qlam = q0 + np.dot(np.transpose(Q), lam)
gvec = np.dot(P, uxinv1) + np.dot(Q, xlinv1)
# Note: NumpPy broadcasting interprets these multiplications as if multiplying by a diagonal matrix!
GG = P * uxinv2.flatten() - Q * xlinv2.flatten()
dpsidx = plam / ux2 - qlam / xl2
delx = dpsidx - epsvecn / (x - alfa) + epsvecn / (beta - x)
dely = self.c + self.d * y - lam - epsvecm / y
delz = self.a0 - np.dot(np.transpose(self.a), lam) - epsi / z
dellam = gvec - self.a * z - y - self.b + epsvecm / lam
diagx = plam / ux3 + qlam / xl3
diagx = 2 * diagx + xsi / (x - alfa) + eta / (beta - x)
diagxinv = np.reciprocal(diagx)
diagy = self.d + mu / y
diagyinv = np.reciprocal(diagy)
diaglam = s / lam
diaglamyi = diaglam + diagyinv
# less constraints then design variables
if m < n:
blam = dellam + dely / diagy - np.dot(
GG, np.divide(delx, diagx))
Alam = np.diagflat(diaglamyi) + np.dot(
GG * diagxinv.flatten(), np.transpose(GG))
# create right hand side
bb = np.transpose(
np.concatenate((np.transpose(blam), delz), axis=1))
# create left hand side
AA = np.concatenate(
(np.concatenate((Alam, self.a), axis=1),
np.concatenate((np.transpose(self.a), -zet / z),
axis=1)))
# solve system
solut = np.linalg.solve(AA, bb)
# store solution
dlam = solut[0:m]
dz = np.array([solut[m]])
dx = -delx / diagx - (np.dot(np.transpose(GG),
dlam)) / diagx
# more constraints then design variables
else:
diaglamyiinv = np.reciprocal(diaglamyi)
dellamyi = dellam + dely / diagy
# prepare LHS
Axx = np.diagflat(diagx) + np.dot(
np.transpose(GG), np.dot(np.diagflat(diaglamyiinv),
GG))
azz = zet / z + np.transpose(self.a) * (a / diaglamyi)
axz = -np.transose(GG) * (a / diaglamyi)
# prepare RHS
bx = delx + np.transose(GG) * (dellamyi / diaglamyi)
bz = delz - np.transose(a) * (dellamyi / diaglamyi)
# create LHS
AA = np.vstack((np.hstack(
(Axx, axz)), np.hstack((np.transpose(axz), azz))))
# create RHS
bb = np.transpose(np.hstack((-np.transpose(bx), -bz)))
# solve system
solut = numpy.linalg.solve(AA, bb)
# store solution
dx = solut[1:n]
dz = np.array([solut[n + 1]])
dlam = (GG * dx) / diaglamyi - dz * (
a / diaglamyi) + dellamyi / diaglamyi
dy = -dely / diagy + dlam / diagy
dxsi = -xsi + epsvecn / (x - alfa) - (xsi * dx) / (x - alfa)
deta = -eta + epsvecn / (beta - x) + (eta * dx) / (beta - x)
dmu = -mu + epsvecm / y - (mu * dy) / y
dzet = -zet + epsi / z - zet * dz / z
ds = -s + epsvecm / lam - (s * dlam) / lam
# construct xx
xx = np.concatenate((y, z, lam, xsi, eta, mu, zet, s))
# construct dxx
dxx = np.concatenate((dy, dz, dlam, dxsi, deta, dmu, dzet, ds))
stepxx = -1.01 * dxx / xx
stmxx = np.amax(stepxx)
stepalfa = -1.01 * dx / (x - alfa)
stmalfa = np.amax(stepalfa)
stepbeta = 1.01 * dx / (beta - x)
stmbeta = np.amax(stepbeta)
stmalbe = np.maximum(stmalfa, stmbeta)
stmalbexx = np.maximum(stmalbe, stmxx)
stminv = np.maximum(stmalbexx, 1)
steg = 1 / stminv
xold = x
yold = y
zold = z
lamold = lam
xsiold = xsi
etaold = eta
muold = mu
zetold = zet
sold = s
# init loop
itto = 0
resinew = 2 * residunorm
while (resinew > residunorm and itto < 50):
# increment loop
itto += 1
x = xold + steg * dx
y = yold + steg * dy
z = zold + steg * dz
lam = lamold + steg * dlam
xsi = xsiold + steg * dxsi
eta = etaold + steg * deta
mu = muold + steg * dmu
zet = zetold + steg * dzet
s = sold + steg * ds
ux1 = self.upp - x
xl1 = x - self.low
ux2 = ux1 * ux1
xl2 = xl1 * xl1
uxinv1 = np.reciprocal(ux1)
xlinv1 = np.reciprocal(xl1)
plam = p0 + np.dot(np.transpose(P), lam)
qlam = q0 + np.dot(np.transpose(Q), lam)
gvec = np.dot(P, uxinv1) + np.dot(Q, xlinv1)
dpsidx = plam / ux2 - qlam / xl2
rex = dpsidx - xsi + eta
rey = self.c + self.d * y - mu - lam
rez = self.a0 - zet - np.transpose(self.a) * lam
relam = gvec - self.a * z - y + s - self.b
rexsi = xsi * (x - alfa) - epsvecn
reeta = eta * (beta - x) - epsvecn
remu = mu * y - epsvecm
rezet = zet * z - epsi
res = lam * s - epsvecm
residu1 = np.concatenate((rex, rey, rez))
residu2 = np.concatenate(
(relam, rexsi, reeta, remu, rezet, res))
# complete residu
residu = np.concatenate((residu1, residu2))
#compute norm
resinew = np.sqrt((residu * residu).sum())
steg = steg / 2
# after while loop add new residu's
residunorm = resinew
residumax = np.max(np.absolute(residu))
steg = 2 * steg
#==============================================================================
# if ittt > 99:
# print("Max iterations in subsolve reached: ittt:", ittt, ", epsi:", epsi)
#==============================================================================
epsi = 0.1 * epsi
# print("MMA: {0} inner iterations.".format(itera))
# After all loops return values
return x, y, z, lam, xsi, eta, mu, zet, s
script_dir = os.path.dirname(__file__)
rel_path = "images/gray.png"
abs_file_path = os.path.join(script_dir, rel_path)
rawImage = PIL.Image.open(abs_file_path)
rawImage.load()
A = np.asarray(rawImage)
Aherm = A.transpose() # Always real-valued, so transpose is fine
AstarA = np.matmul(Aherm, A)
AstarA_eigenValues, AstarA_eigenVectors = np.linalg.eig(AstarA)
#for evalue in AstarA_eigenValues:
# if evalue < 0:
# print(evalue)
U, s, V = np.linalg.svd(A)
S = np.diagflat(s)
Areconstructed = np.matmul(np.matmul(U, S), V).astype(int)
# Show comparison
print(A)
print(Areconstructed)
std = np.std(A - Areconstructed)
print("MSE:")
print(std**2)
l_min=1e-5,
N_max=2000,
sigma=None,
layer=128,
discount_factor=0.99,
scale_reward=False,
mean=True)
render = True
scale = True
theta = np.random.uniform(low=0,
high=1,
size=state_dim + actions_dim + 1 + state_dim)
noise_prior = np.sqrt(theta[-state_dim:])
noise_prior = np.diagflat(noise_prior)
theta = theta[0:-state_dim]
keps = 1.5e-4
B = np.eye(state_dim)
max_episode = 6000
if alpha < 1:
max_episode = 3000
max_step = 300
eval_step = 300
num_test_per_episode = 3
solved = 0
solve_ep = 0
reward_his = []
window_reward_his = deque(maxlen=100)
def write_spectrum(self, row, col, states, meas, geom, flush_immediately=False):
"""Write data from a single inversion to all output buffers."""
self.writes = self.writes + 1
if len(states) == 0:
# Write a bad data flag
atm_bad = np.zeros(len(self.fm.n_chan)*5) * -9999.0
state_bad = np.zeros(len(self.fm.statevec)) * -9999.0
data_bad = np.zeros(self.fm.instrument.n_chan) * -9999.0
to_write = {
'estimated_state_file': state_bad,
'estimated_reflectance_file': data_bad,
'estimated_emission_file': data_bad,
'modeled_radiance_file': data_bad,
'apparent_reflectance_file': data_bad,
'path_radiance_file': data_bad,
'simulated_measurement_file': data_bad,
'algebraic_inverse_file': data_bad,
'atmospheric_coefficients_file': atm_bad,
'radiometry_correction_file': data_bad,
'spectral_calibration_file': data_bad,
'posterior_uncertainty_file': state_bad
}
else:
# The inversion returns a list of states, which are
# intepreted either as samples from the posterior (MCMC case)
# or as a gradient descent trajectory (standard case). For
# gradient descent the last spectrum is the converged solution.
if self.iv.mode == 'mcmc':
state_est = states.mean(axis=0)
else:
state_est = states[-1, :]
# Spectral calibration
wl, fwhm = self.fm.calibration(state_est)
cal = np.column_stack(
[np.arange(0, len(wl)), wl / 1000.0, fwhm / 1000.0])
# If there is no actual measurement, we use the simulated version
# in subsequent calculations. Naturally in these cases we're
# mostly just interested in the simulation result.
if meas is None:
meas = self.fm.calc_rdn(state_est, geom)
# Rodgers diagnostics
lamb_est, meas_est, path_est, S_hat, K, G = \
self.iv.forward_uncertainty(state_est, meas, geom)
# Simulation with noise
meas_sim = self.fm.instrument.simulate_measurement(meas_est, geom)
# Algebraic inverse and atmospheric optical coefficients
x_surface, x_RT, x_instrument = self.fm.unpack(state_est)
rfl_alg_opt, Ls, coeffs = invert_algebraic(self.fm.surface,
self.fm.RT, self.fm.instrument, x_surface, x_RT, x_instrument,
meas, geom)
rhoatm, sphalb, transm, solar_irr, coszen, transup = coeffs
L_atm = self.fm.RT.get_L_atm(x_RT, geom)
L_down_transmitted = self.fm.RT.get_L_down_transmitted(x_RT, geom)
atm = np.column_stack(list(coeffs[:4]) +
[np.ones((len(wl), 1)) * coszen])
atm = atm.T.reshape((len(wl)*5,))
# Upward emission & glint and apparent reflectance
Ls_est = self.fm.calc_Ls(state_est, geom)
apparent_rfl_est = lamb_est + Ls_est
# Radiometric calibration
factors = np.ones(len(wl))
if 'radiometry_correction_file' in self.outfiles:
if 'reference_reflectance_file' in self.infiles:
reference_file = self.infiles['reference_reflectance_file']
self.rfl_ref = reference_file.read_spectrum(row, col)
self.wl_ref = reference_file.wl
w, fw = self.fm.instrument.calibration(x_instrument)
resamp = resample_spectrum(self.rfl_ref, self.wl_ref,
w, fw, fill=True)
meas_est = self.fm.calc_meas(state_est, geom, rfl=resamp)
factors = meas_est / meas
else:
logging.warning('No reflectance reference')
# Assemble all output products
to_write = {
'estimated_state_file': state_est,
'estimated_reflectance_file': np.column_stack((self.fm.surface.wl, lamb_est)),
'estimated_emission_file': np.column_stack((self.fm.surface.wl, Ls_est)),
'modeled_radiance_file': np.column_stack((wl, meas_est)),
'apparent_reflectance_file': np.column_stack((self.fm.surface.wl, apparent_rfl_est)),
'path_radiance_file': np.column_stack((wl, path_est)),
'simulated_measurement_file': np.column_stack((wl, meas_sim)),
'algebraic_inverse_file': np.column_stack((self.fm.surface.wl, rfl_alg_opt)),
'atmospheric_coefficients_file': atm,
'radiometry_correction_file': factors,
'spectral_calibration_file': cal,
'posterior_uncertainty_file': np.sqrt(np.diag(S_hat))
}
for product in self.outfiles:
logging.debug('IO: Writing '+product)
self.outfiles[product].write_spectrum(row, col, to_write[product])
if (self.writes % flush_rate) == 0 or flush_immediately:
self.outfiles[product].flush_buffers()
# Special case! samples file is matlab format.
if self.output.mcmc_samples_file is not None:
logging.debug('IO: Writing mcmc_samples_file')
mdict = {'samples': states}
scipy.io.savemat(self.output.mcmc_samples_file, mdict)
# Special case! Data dump file is matlab format.
if self.output.data_dump_file is not None:
logging.debug('IO: Writing data_dump_file')
x = state_est
xall = states
Seps_inv, Seps_inv_sqrt = self.iv.calc_Seps(x, meas, geom)
meas_est_window = meas_est[self.iv.winidx]
meas_window = meas[self.iv.winidx]
xa, Sa, Sa_inv, Sa_inv_sqrt = self.iv.calc_prior(x, geom)
prior_resid = (x - xa).dot(Sa_inv_sqrt)
rdn_est = self.fm.calc_rdn(x, geom)
rdn_est_all = np.array([self.fm.calc_rdn(xtemp, geom)
for xtemp in states])
x_surface, x_RT, x_instrument = self.fm.unpack(x)
Kb = self.fm.Kb(x, geom)
xinit = invert_simple(self.fm, meas, geom)
Sy = self.fm.instrument.Sy(meas, geom)
cost_jac_prior = np.diagflat(x - xa).dot(Sa_inv_sqrt)
cost_jac_meas = Seps_inv_sqrt.dot(K[self.iv.winidx, :])
meas_Cov = self.fm.Seps(x, meas, geom)
lamb_est, meas_est, path_est, S_hat, K, G = \
self.iv.forward_uncertainty(state_est, meas, geom)
A = np.matmul(K, G)
# Form the MATLAB dictionary object and write to file
mdict = {
'K': K,
'G': G,
'S_hat': S_hat,
'prior_mu': xa,
'Ls': Ls,
'prior_Cov': Sa,
'meas': meas,
'rdn_est': rdn_est,
'rdn_est_all': rdn_est_all,
'x': x,
'xall': xall,
'x_surface': x_surface,
'x_RT': x_RT,
'x_instrument': x_instrument,
'meas_Cov': meas_Cov,
'wl': wl,
'fwhm': fwhm,
'lamb_est': lamb_est,
'coszen': coszen,
'cost_jac_prior': cost_jac_prior,
'Kb': Kb,
'A': A,
'cost_jac_meas': cost_jac_meas,
'winidx': self.iv.winidx,
'windows': self.iv.windows,
'prior_resid': prior_resid,
'noise_Cov': Sy,
'xinit': xinit,
'rhoatm': rhoatm,
'sphalb': sphalb,
'transm': transm,
'transup': transup,
'solar_irr': solar_irr,
'L_atm': L_atm,
'L_down_transmitted': L_down_transmitted
}
scipy.io.savemat(self.output.data_dump_file, mdict)
# Write plots, if needed
if len(states) > 0 and self.output.plot_directory is not None:
if 'reference_reflectance_file' in self.infiles:
reference_file = self.infiles['reference_reflectance_file']
self.rfl_ref = reference_file.read_spectrum(row, col)
self.wl_ref = reference_file.wl
for i, x in enumerate(states):
# Calculate intermediate solutions
lamb_est, meas_est, path_est, S_hat, K, G = \
self.iv.forward_uncertainty(state_est, meas, geom)
plt.cla()
red = [0.7, 0.2, 0.2]
wl, fwhm = self.fm.calibration(x)
xmin, xmax = min(wl), max(wl)
fig = plt.subplots(1, 2, figsize=(10, 5))
plt.subplot(1, 2, 1)
meas_est = self.fm.calc_meas(x, geom)
for lo, hi in self.iv.windows:
idx = np.where(np.logical_and(wl > lo, wl < hi))[0]
p1 = plt.plot(wl[idx], meas[idx], color=red, linewidth=2)
p2 = plt.plot(wl, meas_est, color='k', linewidth=1)
plt.title("Radiance")
plt.title("Measurement (Scaled DN)")
ymax = max(meas)*1.25
ymax = max(meas)+0.01
ymin = min(meas)-0.01
plt.text(500, ymax*0.92, "Measured", color=red)
plt.text(500, ymax*0.86, "Model", color='k')
plt.ylabel(r"$\mu$W nm$^{-1}$ sr$^{-1}$ cm$^{-2}$")
plt.ylabel("Intensity")
plt.xlabel("Wavelength (nm)")
plt.ylim([-0.001, ymax])
plt.ylim([ymin, ymax])
plt.xlim([xmin, xmax])
plt.subplot(1, 2, 2)
lamb_est = self.fm.calc_lamb(x, geom)
ymax = min(max(lamb_est)*1.25, 0.10)
for lo, hi in self.iv.windows:
# black line
idx = np.where(np.logical_and(wl > lo, wl < hi))[0]
p2 = plt.plot(wl[idx], lamb_est[idx], 'k', linewidth=2)
ymax = max(max(lamb_est[idx]*1.2), ymax)
# red line
if 'reference_reflectance_file' in self.infiles:
idx = np.where(np.logical_and(
self.wl_ref > lo, self.wl_ref < hi))[0]
p1 = plt.plot(self.wl_ref[idx], self.rfl_ref[idx],
color=red, linewidth=2)
ymax = max(max(self.rfl_ref[idx]*1.2), ymax)
# green and blue lines - surface components
if hasattr(self.fm.surface, 'components') and \
self.output.plot_surface_components:
idx = np.where(np.logical_and(self.fm.surface.wl > lo,
self.fm.surface.wl < hi))[0]
p3 = plt.plot(self.fm.surface.wl[idx],
self.fm.xa(x, geom)[idx], 'b', linewidth=2)
for j in range(len(self.fm.surface.components)):
z = self.fm.surface.norm(
lamb_est[self.fm.surface.idx_ref])
mu = self.fm.surface.components[j][0] * z
plt.plot(self.fm.surface.wl[idx], mu[idx], 'g:',
linewidth=1)
plt.text(500, ymax*0.86, "Remote estimate", color='k')
if 'reference_reflectance_file' in self.infiles:
plt.text(500, ymax*0.92, "In situ reference", color=red)
if hasattr(self.fm.surface, 'components') and \
self.output.plot_surface_components:
plt.text(500, ymax*0.80, "Prior mean state ",
color='b')
plt.text(500, ymax*0.74, "Surface components ",
color='g')
plt.ylim([-0.0010, ymax])
plt.xlim([xmin, xmax])
plt.title("Reflectance")
plt.title("Source Model")
plt.xlabel("Wavelength (nm)")
fn = os.path.join(self.output.plot_directory, ('frame_%i.png' % i))
plt.savefig(fn)
plt.close()
def initialize_from_hdf5_file(file_name,
structure,
read_trajectory=True,
initial_cut=1,
final_cut=None,
memmap=False):
import h5py
print("Reading data from hdf5 file: " + file_name)
trajectory = None
velocity = None
vc = None
reduced_q_vector = None
#Check file exists
if not os.path.isfile(file_name):
print(file_name + ' file does not exist!')
exit()
hdf5_file = h5py.File(file_name, "r")
if "trajectory" in hdf5_file and read_trajectory is True:
trajectory = hdf5_file['trajectory'][:]
if final_cut is not None:
trajectory = trajectory[initial_cut - 1:final_cut]
else:
trajectory = trajectory[initial_cut - 1:]
if "velocity" in hdf5_file:
velocity = hdf5_file['velocity'][:]
if final_cut is not None:
velocity = velocity[initial_cut - 1:final_cut]
else:
velocity = velocity[initial_cut - 1:]
if "vc" in hdf5_file:
vc = hdf5_file['vc'][:]
if final_cut is not None:
vc = vc[initial_cut - 1:final_cut]
else:
vc = vc[initial_cut - 1:]
if "reduced_q_vector" in hdf5_file:
reduced_q_vector = hdf5_file['reduced_q_vector'][:]
print("Load trajectory projected onto {0}".format(reduced_q_vector))
time = hdf5_file['time'][:]
supercell = hdf5_file['super_cell'][:]
hdf5_file.close()
if vc is None:
return dyn.Dynamics(structure=structure,
trajectory=trajectory,
velocity=velocity,
time=time,
supercell=np.dot(np.diagflat(supercell),
structure.get_cell()),
memmap=memmap)
else:
return vc, reduced_q_vector, dyn.Dynamics(structure=structure,
time=time,
supercell=np.dot(
np.diagflat(supercell),
structure.get_cell()),
memmap=memmap)
def generate_test_trajectory(
structure,
supercell=(1, 1, 1),
minimum_frequency=0.1, # THz
total_time=2, # picoseconds
time_step=0.002, # picoseconds
temperature=400, # Kelvin
silent=False,
memmap=False,
phase_0=0.0):
import random
from dynaphopy.power_spectrum import _progress_bar
print('Generating ideal harmonic data for testing')
kb_boltzmann = 0.831446 # u * A^2 / ( ps^2 * K )
number_of_unit_cells_phonopy = np.prod(
np.diag(structure.get_supercell_phonon()))
number_of_unit_cells = np.prod(supercell)
# atoms_relation = float(number_of_unit_cells)/ number_of_unit_cells_phonopy
#Recover dump trajectory from file (test only)
import pickle
if False:
dump_file = open("trajectory.save", "r")
trajectory = pickle.load(dump_file)
return trajectory
number_of_atoms = structure.get_number_of_cell_atoms()
number_of_primitive_atoms = structure.get_number_of_primitive_atoms()
number_of_dimensions = structure.get_number_of_dimensions()
positions = structure.get_positions(supercell=supercell)
masses = structure.get_masses(supercell=supercell)
number_of_atoms = number_of_atoms * number_of_unit_cells
number_of_primitive_cells = number_of_atoms / number_of_primitive_atoms
atom_type = structure.get_atom_type_index(supercell=supercell)
#Generate additional wave vectors sample
# structure.set_supercell_phonon_renormalized(np.diag(supercell))
q_vector_list = pho_interface.get_commensurate_points(
structure, np.diag(supercell))
q_vector_list_cart = [
np.dot(q_vector,
2 * np.pi * np.linalg.inv(structure.get_primitive_cell()).T)
for q_vector in q_vector_list
]
atoms_relation = float(
len(q_vector_list) * number_of_primitive_atoms) / number_of_atoms
#Generate frequencies and eigenvectors for the testing wave vector samples
print('Wave vectors included in test (commensurate points)')
eigenvectors_r = []
frequencies_r = []
for i in range(len(q_vector_list)):
print(q_vector_list[i])
eigenvectors, frequencies = pho_interface.obtain_eigenvectors_and_frequencies(
structure, q_vector_list[i])
eigenvectors_r.append(eigenvectors)
frequencies_r.append(frequencies)
number_of_frequencies = len(frequencies_r[0])
#Generating trajectory
if not silent:
_progress_bar(0, 'generating')
#Generating trajectory
trajectory = []
for time in np.arange(total_time, step=time_step):
coordinates = np.array(positions[:, :], dtype=complex)
for i_freq in range(number_of_frequencies):
for i_long, q_vector in enumerate(q_vector_list_cart):
if abs(
frequencies_r[i_long][i_freq]
) > minimum_frequency: # Prevent error due to small frequencies
amplitude = np.sqrt(
number_of_dimensions * kb_boltzmann * temperature /
number_of_primitive_cells * atoms_relation) / (
frequencies_r[i_long][i_freq] * 2 * np.pi
) # + random.uniform(-1,1)*0.05
normal_mode = amplitude * np.exp(
np.complex(0, -1) * frequencies_r[i_long][i_freq] *
2.0 * np.pi * time)
phase = np.exp(
np.complex(0, 1) * np.dot(q_vector, positions.T) +
phase_0)
coordinates += (1.0 / np.sqrt(masses)[None].T *
eigenvectors_r[i_long][i_freq, atom_type] *
phase[None].T * normal_mode).real
trajectory.append(coordinates)
if not silent:
_progress_bar(
float(time + time_step) / total_time,
'generating',
)
trajectory = np.array(trajectory)
time = np.array([i * time_step for i in range(trajectory.shape[0])],
dtype=float)
energy = np.array([
number_of_atoms * number_of_dimensions * kb_boltzmann * temperature
for i in range(trajectory.shape[0])
],
dtype=float)
#Save a trajectory object to file for later recovery (test only)
if False:
dump_file = open("trajectory.save", "w")
pickle.dump(
dyn.Dynamics(structure=structure,
trajectory=np.array(trajectory, dtype=complex),
energy=np.array(energy),
time=time,
supercell=np.dot(np.diagflat(supercell),
structure.get_cell())), dump_file)
dump_file.close()
# structure.set_supercell_phonon_renormalized(None)
return dyn.Dynamics(structure=structure,
trajectory=np.array(trajectory, dtype=complex),
energy=np.array(energy),
time=time,
supercell=np.dot(np.diagflat(supercell),
structure.get_cell()),
memmap=memmap)
])
A3 = np.array([
[1.0, 0, DiscTimePeriod, 0],
[0, 1.0, 0, DiscTimePeriod],
[
0, 0,
np.cos(RotRate * DiscTimePeriod),
1 * np.sin(RotRate * DiscTimePeriod)
],
[
0, 0, -1 * np.sin(RotRate * DiscTimePeriod),
np.cos(RotRate * DiscTimePeriod)
],
])
V1 = np.diagflat([0.0025, 0.0025, 0.0005, 0.0005])
# W1 = np.diagflat([0.000025, 0.000025])
W1 = np.diagflat([0.0005, 0.0005])
WL1 = np.diagflat([0.32, 0.32])
# From https://docs.px4.io/master/en/advanced_config/tuning_the_ecl_ekf.html
# and https://dewesoft.com/products/interfaces-and-sensors/gps-and-imu-devices/tech-specs
C1 = np.array([[0, 0, 1, 0], [0, 0, 0, 1]])
CL1 = np.array([[1, 0, 0, 0], [0, 1, 0, 0]])
D = np.array([[1, 0], [0, 1]])
A = TimeVaryingValue([A1, A2, A3], modes)
V = TimeVaryingValue([V1, V1, V1], modes)
def Dsoftmax(x):
s = x.reshape(-1, 1)
return np.diagflat(s) - np.dot(s, s.T)
def time_diagflat_l100(self):
np.diagflat(self.l100)
def fractionalDifferenciatorWeights(p,
alpha,
NbPoles=20,
PolesMinMax=(-5, 10),
NbFreqPoints=200,
FreqsMinMax=(1, 48e3),
DoPlot=True):
# Defintion of the frequency grid
fmin, fmax = FreqsMinMax
wmin, wmax = 2 * np.pi * fmin, 2 * np.pi * fmax
w = np.exp(
np.log(wmin) +
np.linspace(0, 1, NbFreqPoints + 1) * np.log(wmax / wmin))
w12 = np.sqrt(w[1:] * w[:-1])
# Unpack min and max exponents to define the list of poles
emin, emax = PolesMinMax
Xi = np.logspace(emin, emax, NbPoles) # xi_0 -> xi_{N+1}
# Input to Output transfer function of the fractional integrator of
# order 1-alpha
beta = 1. - alpha
def transferFunctionFracInt(s):
return s**-beta
# Target transfer function evaluated on the frequency grid
T = transferFunctionFracInt(1j * w12)
# Return the basis vector of elementary damping with poles Xi
def Basis(s, Xi):
return (s + Xi)**-1
# Matrix of basis transfer function for each poles on the frequency grid
M = np.zeros((NbFreqPoints, NbPoles), dtype=np.complex64)
for k in np.arange(NbFreqPoints):
M[k, :] = Basis(1j * w12[k], Xi)
# Perceptual weights
WBuildingVector = (np.log(w[1:]) - np.log(w[:-1])) / (np.abs(T)**2)
W = np.diagflat(WBuildingVector)
# Definition of the cost function
def CostFunction(mu):
mat = np.dot(M, mu) - T
cost = np.dot(np.conjugate(mat.T), np.dot(W, mat))
return cost.real
# Optimization constraints
bnds = [(0, None) for n in range(NbPoles)]
# Optimization
from scipy.optimize import minimize
MuOpt = minimize(CostFunction, np.ones(NbPoles), bounds=bnds, tol=EPS)
Mu = MuOpt.x # Get the solution
# Conversion to phs parameters
diagQ = []
diagR = []
# Eliminate 0 valued weigths
for n in np.arange(NbPoles):
if Mu[n] > 0:
diagR.append(p * Mu[n])
diagQ.append(p * Mu[n] * Xi[n])
if DoPlot:
from matplotlib.pyplot import (figure, subplot, plot, semilogx, ylabel,
legend, grid, xlabel)
TOpt = np.array(M * np.matrix(Mu).T)
wmin, wmax = 2 * np.pi * fmin, 2 * np.pi * fmax
figure()
subplot(2, 1, 1)
faxis = w12[(wmin < w12) & (w12 < wmax)] / (2 * np.pi)
v1 = 20 * np.log10(np.abs(T[(wmin < w12) & (w12 < wmax)]))
v2 = 20 * np.log10(np.abs(TOpt[(wmin < w12) & (w12 < wmax)]))
v3 = list(map(lambda x, y: x - y, v1, v2))
semilogx(faxis, v1, label='Target')
semilogx(faxis, v2, label='Approx')
ylabel('Transfert (dB)')
legend(loc=0)
grid()
subplot(2, 1, 2)
plot(faxis, v3, label='Error')
xlabel('Log-frequencies (log Hz)')
ylabel('Error (dB)')
legend(loc=0)
grid()
return diagR, diagQ
def fit_firth(logit_model,
start_vec,
X,
y,
step_limit=1000,
convergence_limit=0.0001):
"""Do firth regression
Args:
logit (statsmodels.discrete.discrete_model.Logit)
Logistic model
start_vec (numpy.array)
Pre-initialized vector to speed-up convergence (n, 1)
X (numpy.array)
(n, m)
y (numpy.array)
(n, )
step_limit (int)
Maximum number of iterations
convergence_limit (float)
Convergence tolerance
Returns:
intercept (float)
Intercept
kbeta (float)
Variant beta
beta (iterable)
Covariates betas (n-2)
bse (float)
Beta std-err
fitll (float or None)
Likelihood of fit or None if could not fit
"""
beta_iterations = []
beta_iterations.append(start_vec)
for i in range(0, step_limit):
pi = logit_model.predict(beta_iterations[i])
W = np.diagflat(np.multiply(pi, 1 - pi))
var_covar_mat = np.linalg.pinv(
-logit_model.hessian(beta_iterations[i]))
# build hat matrix
rootW = np.sqrt(W)
H = np.dot(np.transpose(X), np.transpose(rootW))
H = np.matmul(var_covar_mat, H)
H = np.matmul(np.dot(rootW, X), H)
# penalised score
U = np.matmul(np.transpose(X),
y - pi + np.multiply(np.diagonal(H), 0.5 - pi))
new_beta = beta_iterations[i] + np.matmul(var_covar_mat, U)
# step halving
j = 0
while firth_likelihood(new_beta, logit_model) > firth_likelihood(
beta_iterations[i], logit_model):
new_beta = beta_iterations[i] + 0.5 * (new_beta -
beta_iterations[i])
j = j + 1
if (j > step_limit):
return None
beta_iterations.append(new_beta)
if i > 0 and (
np.linalg.norm(beta_iterations[i] - beta_iterations[i - 1]) <
convergence_limit):
break
return_fit = None
if np.linalg.norm(beta_iterations[i] -
beta_iterations[i - 1]) >= convergence_limit:
pass
else:
# Calculate stats
fitll = -firth_likelihood(beta_iterations[-1], logit_model)
intercept = beta_iterations[-1][0]
if len(beta_iterations[-1]) > 1:
kbeta = beta_iterations[-1][1]
bse = math.sqrt(-logit_model.hessian(beta_iterations[-1])[1, 1])
else:
# Encountered when fitting null without any distances/covariates
kbeta = None
bse = None
if len(beta_iterations[-1]) > 2:
beta = beta_iterations[-1][2:].tolist()
else:
beta = None
return_fit = intercept, kbeta, beta, bse, fitll
return return_fit
lensmodel,
intrinsics_true[i]) \
for i in range(len(intrinsics_true))]),
0,1)
# Let's define the observation-time pixel noise. The noise vector
# q_true_sampled_noise has the same shape as q_true for each sample. so
# q_true_sampled_noise.shape = (Nsamples,Npoints,Ncameras,2). The covariance is
# a square matrix with each dimension of length Npoints*Ncameras*2
N_q_true_noise = Npoints*args.Ncameras*2
sigma_qt_sq = \
args.pixel_uncertainty_stdev_triangulation * \
args.pixel_uncertainty_stdev_triangulation
var_qt = np.diagflat( (sigma_qt_sq,) * N_q_true_noise )
var_qt_reshaped = var_qt.reshape( Npoints, args.Ncameras, 2,
Npoints, args.Ncameras, 2 )
if args.Ncameras != 2:
raise Exception("Ncameras == 2 is assumed here")
for ipt in range(Npoints):
var_qt_reshaped[ipt,0,0, ipt,1,0] = sigma_qt_sq*args.pixel_uncertainty_triangulation_correlation
var_qt_reshaped[ipt,1,0, ipt,0,0] = sigma_qt_sq*args.pixel_uncertainty_triangulation_correlation
var_qt_reshaped[ipt,0,1, ipt,1,1] = sigma_qt_sq*args.pixel_uncertainty_triangulation_correlation
var_qt_reshaped[ipt,1,1, ipt,0,1] = sigma_qt_sq*args.pixel_uncertainty_triangulation_correlation
# Let's actually apply the noise to compute var(distancep) empirically to compare
# against the var(distancep) prediction I just computed
# shape (Nsamples,Npoints,Ncameras,2)
qt_noise = \
def baro(self, iterative, chainvel0):
def update_baro_vel():
# updates the barostat velocity tensor
if chainvel0 is not None:
# iL v_{xi} v_g h/8
self.vel_press *= np.exp(-self.timestep_press*chainvel0/8)
# definition of P_intV and G
ptens_vol = np.dot(iterative.vel.T*iterative.masses, iterative.vel) - iterative.vtens
ptens_vol = 0.5*(ptens_vol.T + ptens_vol)
G = (ptens_vol+(2.0*iterative.ekin/iterative.ndof-self.press*iterative.ff.system.cell.volume)*np.eye(3))/self.mass_press
if not self.anisotropic:
G = np.trace(G)
if self.vol_constraint:
G -= np.trace(G)/self.dim*np.eye(self.dim)
# iL G_g h/4
self.vel_press += G*self.timestep_press/4
if chainvel0 is not None:
# iL v_{xi} v_g h/8
self.vel_press *= np.exp(-self.timestep_press*chainvel0/8)
# first part of the barostat velocity tensor update
update_baro_vel()
# iL v_g h/2
if self.anisotropic:
Dr, Qg = np.linalg.eigh(self.vel_press)
Daccr = np.diagflat(np.exp(Dr*self.timestep_press/2))
rot_mat = np.dot(np.dot(Qg, Daccr), Qg.T)
pos_new = np.dot(iterative.pos, rot_mat)
rvecs_new = np.dot(iterative.rvecs, rot_mat)
else:
c = np.exp(self.vel_press*self.timestep_press/2)
pos_new = c*iterative.pos
rvecs_new = c*iterative.rvecs
# update the positions and cell vectors
iterative.ff.update_pos(pos_new)
iterative.pos[:] = pos_new
iterative.ff.update_rvecs(rvecs_new)
iterative.rvecs[:] = rvecs_new
# update the potential energy
iterative.gpos[:] = 0.0
iterative.vtens[:] = 0.0
iterative.epot = iterative.ff.compute(iterative.gpos,iterative.vtens)
# -iL (v_g + Tr(v_g)/ndof) h/2
if self.anisotropic:
if self.vol_constraint:
Dg, Eg = np.linalg.eigh(self.vel_press)
else:
Dg, Eg = np.linalg.eigh(self.vel_press+(np.trace(self.vel_press)/iterative.ndof)*np.eye(3))
Daccg = np.diagflat(np.exp(-Dg*self.timestep_press/2))
rot_mat = np.dot(np.dot(Eg, Daccg), Eg.T)
vel_new = np.dot(iterative.vel, rot_mat)
else:
vel_new = np.exp(-((1.0+3.0/iterative.ndof)*self.vel_press)*self.timestep_press/2) * iterative.vel
# update the velocities and the kinetic energy
iterative.vel[:] = vel_new
iterative.ekin = iterative._compute_ekin()
# second part of the barostat velocity tensor update
update_baro_vel()