def poisson(tri, boundary):
"""
Get elements and vertices from meshgrid using the triangle tri and boundary points boundary
"""
elements = tri.triangle_nodes
vertices = np.vstack((tri.x, tri.y)).T
# number of vertices and elements
N = vertices.shape[0]
E = elements.shape[0]
# Loop over elements and assemble LHS and RHS
A = np.matlib.zeros((N, N))
b = np.matlib.zeros((N, 1))
for j in range(E):
index = (elements[j, :]).tolist()
A[ix_(index, index)] += A_e(vertices[index, :])
b[index] += b_e(vertices[index, :])
# find the "free" vertices that we need to solve for
free = list(set(range(len(vertices))) - set(boundary))
# initialise solution to zero so "non-free" vertices are by default zero
u = np.matlib.zeros((N, 1))
# solve for "free" vertices.
u[free] = np.solve(A[ix_(free, free)], b[free])
return np.array(u)
def getRecessionParams(discharge):
dQdt, avgQ = getDerivative(discharge)
y = log(-dQdt)
A = np.concatenate((np.ones(len(discharge) - 1, 1), math.log(avgQ)),
axis=0)
x = np.solve(A, y)
a = math.exp(x[0])
b = x[1]
return a, b
def evaluate_policy(self, P, R, policy):
P_pi = np.zeros((self.nS, self.nS))
R_pi = np.zeros((self.nS, self.nA))
for s in range(self.nS):
for s1 in range(self.nS):
P_pi = np.sum(policy[s, :] * P[s, :, s1])
for a in range(self.nA):
R_pi = policy[s, a] * np.sum(P[s, a, :] * R[s, a, :])
I = np.diag(np.ones(self.nS))
V = np.solve(I - self.mdp_info.gamma * P_pi, R_pi)
return V
def _solve_(self):
Z = self.P_Omega.toarray()
Known = self.Omega
if self.init == 0:
X = np.zeros((self.n1, self.k))
Y = np.eye(self.k, self.n2)
Res = self.M_Omega
res = self.datanrm
if self.est_rank == 1: rank_max = min(self.rank_max, self.k)
#if self.n1>self.n2: Z=Z.T
#parameter for alf
alf, increment, itr_rank = 0, 1, 0
RMSE = []
while True:
X0, Y0, Res0, res0 = X, Y, Res, res
itr_rank += 1
if self.Zfull:
Z0 = Z
X = np.dot(Z, Y.T)
if self.est_rank == 1:
X, R = np.linalg.qr(X)
Y = np.dot(X.T, Z)
elif self.DoQR:
X, R = np.linalg.qr(X)
Y = np.dot(X.T, Z)
else:
Xt = X.T
Y = np.solve(np.dot(Xt, X), np, dot(Xt, Z))
Z = np.dot(X, Y)
Res = self.M_Omega - Z[Known]
res = np.linalg.norm(Res, 2)
relres = res / self.datanrm
ratio = res / res0
reschg = abs(1 - res / res0)
RMSE.append(res / np.linalg.norm(self.Morigin, ord='fro'))
if itr_rank == 100: break
x_coordinate = range(len(RMSE))
plt.title('GMM-NOISE')
plt.xlabel('Number of iterations')
plt.ylabel('RMSE')
#lg scale
#plt.yscale('log')
plt.plot(x_coordinate, RMSE, '-')
plt.show()
return 0
def gauss_newton(x, y, r, p0):
length = len(x)
initial = np.array(p0)
x_solve = [length]
y_solve = [length]
vk = [length]
x_solve[0] = p0[0]
y_solve[0] = p0[0]
vk[0] = p0[0]
for i in range(length):
A = fsolve(x_solve[i], p0)
vk[i] = np.solve(
-np.transpose(A) * np.sqrt((p0[0] - x_solve[i])**2 +
(p0[1] - y_solve[i])**2),
np.transpose(A) * A)
x_solve[i + 1] = x_solve[i] + vk[i]
print(x_solve[length])
def simpson(func, v):
a = v[0] #Constantes
b = v[1]
x = Symbol('x') #Inicializa "x" como símbolo
f = sympify(func) #Se traduce el string "func" a una función de sympy
fx = lambdify(x, f, modules=['numpy']) #Se inicializa la función f(x)
x0 = a
x1 = (a + b) / 2
x2 = b
h = (b - a) / 2 #Valor h
f_resultado = (h / 3) * (fx(x0) + 4 * fx(x1) + fx(x2)
) #Resultado de la aproximacion
df = f.diff(x, 4) #Cuarta derivada
dfx = lambdify(x, df, modules=['numpy']) #Se iniacilaiza la funcion f4(x)
max_relativo = maximo(dfx,
v) #Calculo del maximo en los extremos del intervalo
#En casa de que la funcion no se indefina f(x) = 0
try:
sol = np.solve(f.diff(x, 1))
punto = []
for i in sol: #Calcula cual de los resultados es el maximo
if (abs(dfx(i)) > punto):
punto = [i, abs(dfx(i))]
if (punto[1] > max_relativo[1]
): #Compara el maximo con el maximo de los extremos
error = (h**5 / 90) * abs(dfx(punto[0]))
else:
error = (h**5 / 90) * abs(dfx(max_relativo[0]))
except:
error = (h**5 / 90) * abs(dfx(max_relativo[0]))
print(f_resultado, error)
return (f_resultado, error) #Resultado (aproximacion, error)
def test_lu(self):
import math
A = np.array([[3, 17, 10], [2, 4, -2], [6, 18, -12]])
LU = np.array([[6, 18, -12], [1 / 2, 8, 16], [1 / 3, -1 / 4, 6]])
P = np.array([2, 0, 1], dtype=np.uint16)
# lu decomposition
lu, p = np.lu(A)
for i, ei in enumerate(lu):
for j, eij in enumerate(ei):
self.assertEqual(eij, LU[i][j])
for i, ei in enumerate(p):
self.assertEqual(ei, P[i])
# determinant
self.assertEqual(np.det(A), 6 * 8 * 6)
# zero determinant, singular
AA = np.array([[3, 17, 10], [2, 4, -2], [3, 17, 10]])
self.assertEqual(np.det(AA), 0)
# vector solve
b = np.array([1, -1, 0])
x = np.solve(A, b)
X = [-1.375, 0.375, -0.125]
for i, ei in enumerate(x):
self.assertEqual(ei, X[i])
res = np.dot(A, x) - b
for i, ei in enumerate(res):
self.assertEqual(ei, 0)
# vector solve. singular
with self.assertRaises(ValueError):
x = np.solve(AA, b)
# matrix solve
b = np.array([[1, 0], [-1, 1], [0, 0]])
x = np.solve(A, b)
X = [[-1.375, 4 / 3], [0.375, -1 / 3], [-0.125, 1 / 6]]
for i, ei in enumerate(x):
for j, eij in enumerate(ei):
self.assertEqual(eij, X[i][j])
res = np.dot(A, x) - b
for i, ei in enumerate(res):
for j, eij in enumerate(ei):
self.assertTrue(math.fabs(eij) < 1e-6)
# matrix inverse
Ai = np.inv(A)
I = np.eye(3)
res = np.dot(A, Ai) - I
for i, ei in enumerate(res):
for j, eij in enumerate(ei):
self.assertTrue(math.fabs(eij) < 1e-6)
res = np.dot(Ai, A) - I
for i, ei in enumerate(res):
for j, eij in enumerate(ei):
self.assertTrue(math.fabs(eij) < 1e-6)
Ais = np.solve(A, I)
res = Ai - Ais
for i, ei in enumerate(res):
for j, eij in enumerate(ei):
self.assertTrue(math.fabs(eij) < 1e-6)
# inverse. singular
with self.assertRaises(ValueError):
Ai = np.inv(AA)
def mls(B, v, umin, umax, Wv=None, Wu=None, ud=None, u=None, W=None, imax=100):
"""
mls - Control allocation using minimal least squares.
[u,W,iter] = mls_alloc(B,v,umin,umax,[Wv,Wu,ud,u0,W0,imax])
Solves the bounded sequential least-squares problem
min ||Wu(u-ud)|| subj. to u in M
where M is the set of control signals solving
min ||Wv(Bu-v)|| subj. to umin <= u <= umax
using a two stage active set method. Wu must be diagonal since the
problem is reformulated as a minimal least squares problem. The
implementation does not handle the case of coplanar controls.
Inputs:
-------
B control effectiveness matrix (k x m)
v commanded virtual control (k x 1)
umin lower position limits (m x 1)
umax upper position limits (m x 1)
Wv virtual control weighting matrix (k x k) [I]
Wu control weighting matrix (m x m), diagonal [I]
ud desired control (m x 1) [0]
u0 initial point (m x 1)
W0 initial working set (m x 1) [empty]
imax max no. of iterations [100]
Outputs:
-------
u optimal control
W optimal active set
iter no. of iterations (= no. of changes in the working set + 1)
0 if u_i not saturated
Active set syntax: W_i = -1 if u_i = umin_i
+1 if u_i = umax_i
Directly Based on the code from:
Ola Haerkegard, www.control.isy.liu.se/~ola
see licsence.
"""
#k = number of virtual controls
#m = number of variables (actuators)
k, m = B.shape
if u == None:
u = np.mean(umin + umax, 0)[:, None]
if W == None:
W = np.zeros((m, 1))
if ud == None:
ud = np.zeros((m, 1))
if Wu == None:
Wu = np.eye(m)
if Wv == None:
Wv = np.eye(k)
phase = 1
#Reformulate as a minimal least squares problem. See 2002-03-08 (1).
A = Wv.dot(B).dot(np.linalg.pinv(Wu))
b = Wv.dot(v - B.dot(ud))
xmin = (umin - ud).flatten()
xmax = (umax - ud).flatten()
# Compute initial point and residual.
x = Wu.dot(u - ud)
r = np.atleast_2d(A.dot(x) - b)
#Determine indeces of free variables
i_free = (W == 0).flatten()
m_free = np.sum(i_free)
for i in range(imax):
#print 'Iter: ', i
if phase == 1:
A_free = A[:, i_free]
if m_free <= k:
if m_free > 0:
p_free = np.linalg.lstsq(-A_free, r)[0]
else:
q1, r1 = qr(A_free.T)
p_free = -q1.dot(np.solve(r1.T, r))
p = np.zeros((m, 1))
if A.shape[1] > 1:
p[i_free] = p_free
else:
p[i_free] = p_free.flatten()
else:
i_fixed = np.logical_not(i_free)
m_fixed = m - m_free
if m_fixed > 0:
HT = U[i_fixed.squeeze(), :].T
V, Rtot = qr(np.atleast_2d(HT))
V1 = V[:, :m_fixed]
V2 = V[:, m_fixed + 1:]
R = Rtot[:, m_fixed]
else:
V, Rtot = np.array([[]]), np.array([[]])
V1 = V2 = R = V.T
s = -V2.T.dot(z)
pz = V2.dot(s)
p = U.dot(pz)
x_opt = x + p
infeasible = np.logical_or(x_opt < xmin, x_opt > xmax)
if not np.any(infeasible[i_free]):
x = x_opt
if phase == 1:
r = r + A.dot(p)
else:
z = z + pz
if phase == 1 and m_free >= k:
phase = 2
Utot, Stot = qr(A.T)
U = Utot[:, k:]
z = U.T.dot(x)
else:
lam = np.zeros((m, 1))
if m_free < m:
if phase == 1:
g = A.T.dot(r)
lam = -W * g
else:
lam[i_fixed] = -W[i_fixed] * np.linalg.solve(
R, V1.T.dot(z))
if np.all(lam >= -eps):
u = np.linalg.solve(Wu, x) + ud
return u
lambda_neg, i_neg = np.min(lam), np.argmin(lam)
W[i_neg] = 0
i_free[i_neg] = True
m_free += 1
else:
dist = np.ones(m)
i_min = np.logical_and(i_free, p.flat < 0).flatten()
i_max = np.logical_and(i_free, p.flat > 0).flatten()
dist[i_min] = (xmin[i_min] - x[i_min]) / p[i_min]
dist[i_max] = (xmax[i_max] - x[i_max]) / p[i_max]
alpha, i_alpha = np.min(dist), np.argmin(dist)
x = x + alpha * p
if phase == 1:
r = r + A.dot(alpha * p) #!!
else:
z = z + alpha * pz
W[i_alpha] = np.sign(p[i_alpha])
if i_free[i_alpha]:
i_free[i_alpha] = False
m_free -= 1
u = np.linalg.solve(Wu, x) + ud
return u
def mls(B,v,umin,umax,Wv=None,Wu=None,ud=None,u=None,W=None,imax=100):
"""
mls - Control allocation using minimal least squares.
[u,W,iter] = mls_alloc(B,v,umin,umax,[Wv,Wu,ud,u0,W0,imax])
Solves the bounded sequential least-squares problem
min ||Wu(u-ud)|| subj. to u in M
where M is the set of control signals solving
min ||Wv(Bu-v)|| subj. to umin <= u <= umax
using a two stage active set method. Wu must be diagonal since the
problem is reformulated as a minimal least squares problem. The
implementation does not handle the case of coplanar controls.
Inputs:
-------
B control effectiveness matrix (k x m)
v commanded virtual control (k x 1)
umin lower position limits (m x 1)
umax upper position limits (m x 1)
Wv virtual control weighting matrix (k x k) [I]
Wu control weighting matrix (m x m), diagonal [I]
ud desired control (m x 1) [0]
u0 initial point (m x 1)
W0 initial working set (m x 1) [empty]
imax max no. of iterations [100]
Outputs:
-------
u optimal control
W optimal active set
iter no. of iterations (= no. of changes in the working set + 1)
0 if u_i not saturated
Active set syntax: W_i = -1 if u_i = umin_i
+1 if u_i = umax_i
Directly Based on the code from:
Ola Härkegård, www.control.isy.liu.se/~ola
see licsence.
"""
#k = number of virtual controls
#m = number of variables (actuators)
k,m=B.shape
if u==None:
u=np.mean(umin+umax,0)[:,None]
if W==None:
W=np.zeros((m,1))
if ud==None:
ud=np.zeros((m,1))
if Wu==None:
Wu=np.eye(m)
if Wv==None:
Wv=np.eye(k)
phase=1
#Reformulate as a minimal least squares problem. See 2002-03-08 (1).
A=Wv.dot(B).dot(np.linalg.pinv(Wu))
#print B, v
#A=B
b = Wv.dot(v-B.dot(ud))
#b=v
#print b
xmin = (umin-ud)
xmax = (umax-ud)
# Compute initial point and residual.
x = Wu.dot(u-ud)
#x#=umin-umax
r = A.dot(x)-b
# print x.shape, r.shape, b.shape,x,r
#Determine indeces of free variables
i_free = W==0
m_free = np.sum(i_free)
for i in range(imax):
#print 'Iter: ', i
if phase==1:
A_free = A[:,i_free.squeeze()]
if m_free<=k:
if m_free>0:
p_free=np.linalg.lstsq(-A_free,r)[0]
else:
q1,r1=qr(A_free.T)
p_free=-q1.dot(np.solve(r1.T,r))
p=np.zeros((m,1))
p[i_free.squeeze()]=p_free
else:
i_fixed=np.logical_not(i_free)
m_fixed=m-m_free
if m_fixed>0:
HT=U[i_fixed.squeeze(),:].T
V,Rtot= qr(np.atleast_2d(HT))
V1=V[:,:m_fixed]
V2=V[:,m_fixed+1:]
R=Rtot[:,m_fixed]
else:
V,Rtot=np.array([[]]),np.array([[]])
V1=V2=R=V.T
s=-V2.T.dot(z)
pz=V2.dot(s)
p=U.dot(pz)
x_opt=x+p
infeasible=np.logical_or(x_opt<xmin,x_opt>xmax)
if not np.any(infeasible[i_free]):
x=x_opt
if phase==1:
r=r+A.dot(p)
else:
z=z+pz
if phase==1 and m_free>=k:
phase=2
Utot, Stot=qr(A.T)
U=Utot[:,k:]
z=U.T.dot(x)
else:
lam=np.zeros((m,1))
if m_free<m:
if phase==1:
g=A.T.dot(r)
lam=-W*g
else:
lam[i_fixed]=-W[i_fixed]*np.solve(R,V1.T.dot(z))
if np.all(lam>= -eps):
u=np.linalg.solve(Wu,x)+ud
return u
lambda_neg,i_neg=np.min(lam),np.argmin(lam)
W[i_neg]=0
i_free[i_neg]=1
m_free+=1
else:
dist=np.ones((m,1))
i_min=np.logical_and(i_free,p<0)
i_max=np.logical_and(i_free,p>0)
dist[i_min]=(xmin[i_min]-x[i_min])/p[i_min]
dist[i_max]=(xmax[i_max]-x[i_max])/p[i_max]
alpha,i_alpha=np.min(dist),np.argmin(dist)
x = x + alpha*p
if phase==1:
r=r+A.dot(alpha*p) #!!
else:
z=z+alpha*pz
W[i_alpha]=np.sign(p[i_alpha])
i_free[i_alpha]=0
m_free-=1
u=np.linalg.solve(Wu,x)+ud
return u
def main():
parser = OptionParser(usage="usage: %prog [options] INPUT.mat OUTPUT.mat")
if os.name == "posix":
parser.add_option(
"-p",
"--parallel",
dest="parallel",
help="number of threads "
"(default: available CPUs; use 0 to disable threading code)",
default=None,
metavar="N")
parser.add_option(
"-c",
"--chunksize",
dest="chunk_size",
default=DEFAULT_X_STRIP_SIZE,
metavar="N",
help="number of matrix rows to process at a time in each "
"parallel job (higher -> more memory usage, "
"less parallelism, reduced threading overhead)")
options, args = parser.parse_args()
if len(args) != 2:
parser.error("wrong number of arguments")
if os.name != "posix":
options.parallel = 0
(input_path, output_path) = args
input_data = loadmat(input_path, matlab_compatible=True)
# A hack: we stash the big data into a global variable before using
# multiproessing to fork off threads, so the child processes will inherit
# the data in shared memory. (This only works on Unix, which is why we
# disable the multi-threading code on other systems.)
global _GLOBAL_SHARED_DATA_
g = {}
g["eeg_data"] = input_data["eeg_data"]
# -1 converts from MATLAB-style indexing to Python-style indexing
g["event_idx"] = input_data["event_idx"].squeeze() - 1
g["design_matrix"] = input_data["design_matrix"]
g["pre_event_samples"] = input_data["pre_event_samples"].item()
g["post_event_samples"] = input_data["post_event_samples"].item()
g["artifact_starts"] = input_data["artifact_starts"].squeeze() - 1
# Need a -1 to convert this to 0-based indexing, then a +1 to convert it
# to Python-style half-open interval ranges instead of MATLAB-style closed
# interval ranges... which cancel out, so in fact we can use the numbers
# as is:
g["artifact_stops"] = input_data["artifact_stops"].squeeze()
_GLOBAL_SHARED_DATA_ = g
num_channels = g["eeg_data"].shape[1]
epoch_len = g["pre_event_samples"] + 1 + g["post_event_samples"]
X_columns = epoch_len * g["design_matrix"].shape[1]
XtX_accumulator = np.zeros((X_columns, X_columns))
XtY_accumulator = np.zeros((X_columns, num_channels))
if options.parallel == 0:
# Serial code, in-process
pool = None
imap_fn = itertools.imap
else:
# With any other value for parallel, we spawn worker processes. (This
# includes parallel=1. parallel=1 won't be any faster than parallel=0,
# but it might be useful for testing the worker process code.)
pool = multiprocessing.Pool(options.parallel)
imap_fn = pool.imap_unordered
try:
for (XtX, XtY) in imap_fn(
compute_XtX_XtY_for_slice,
pick_slices(g["eeg_data"].shape[0], g["artifact_starts"],
g["artifact_stops"], options.chunksize)):
XtX_accumulator += XtX
XtY_accumulator += XtY
finally:
if pool is not None:
pool.terminate()
betas = np.solve(XtX_accumulator, XtY_accumulator)
savemat(output_path, {"betas": betas}, oned_as="column")
def main():
parser = OptionParser(usage="usage: %prog [options] INPUT.mat OUTPUT.mat")
if os.name == "posix":
parser.add_option("-p", "--parallel", dest="parallel",
help="number of threads "
"(default: available CPUs; use 0 to disable threading code)",
default=None,
metavar="N")
parser.add_option("-c", "--chunksize", dest="chunk_size",
default=DEFAULT_X_STRIP_SIZE,
metavar="N",
help="number of matrix rows to process at a time in each "
"parallel job (higher -> more memory usage, "
"less parallelism, reduced threading overhead)")
options, args = parser.parse_args()
if len(args) != 2:
parser.error("wrong number of arguments")
if os.name != "posix":
options.parallel = 0
(input_path, output_path) = args
input_data = loadmat(input_path, matlab_compatible=True)
# A hack: we stash the big data into a global variable before using
# multiproessing to fork off threads, so the child processes will inherit
# the data in shared memory. (This only works on Unix, which is why we
# disable the multi-threading code on other systems.)
global _GLOBAL_SHARED_DATA_
g = {}
g["eeg_data"] = input_data["eeg_data"]
# -1 converts from MATLAB-style indexing to Python-style indexing
g["event_idx"] = input_data["event_idx"].squeeze() - 1
g["design_matrix"] = input_data["design_matrix"]
g["pre_event_samples"] = input_data["pre_event_samples"].item()
g["post_event_samples"] = input_data["post_event_samples"].item()
g["artifact_starts"] = input_data["artifact_starts"].squeeze() - 1
# Need a -1 to convert this to 0-based indexing, then a +1 to convert it
# to Python-style half-open interval ranges instead of MATLAB-style closed
# interval ranges... which cancel out, so in fact we can use the numbers
# as is:
g["artifact_stops"] = input_data["artifact_stops"].squeeze()
_GLOBAL_SHARED_DATA_ = g
num_channels = g["eeg_data"].shape[1]
epoch_len = g["pre_event_samples"] + 1 + g["post_event_samples"]
X_columns = epoch_len * g["design_matrix"].shape[1]
XtX_accumulator = np.zeros((X_columns, X_columns))
XtY_accumulator = np.zeros((X_columns, num_channels))
if options.parallel == 0:
# Serial code, in-process
pool = None
imap_fn = itertools.imap
else:
# With any other value for parallel, we spawn worker processes. (This
# includes parallel=1. parallel=1 won't be any faster than parallel=0,
# but it might be useful for testing the worker process code.)
pool = multiprocessing.Pool(options.parallel)
imap_fn = pool.imap_unordered
try:
for (XtX, XtY) in imap_fn(compute_XtX_XtY_for_slice,
pick_slices(g["eeg_data"].shape[0],
g["artifact_starts"],
g["artifact_stops"],
options.chunksize)):
XtX_accumulator += XtX
XtY_accumulator += XtY
finally:
if pool is not None:
pool.terminate()
betas = np.solve(XtX_accumulator, XtY_accumulator)
savemat(output_path, {"betas": betas}, oned_as="column")
