def fit_polynomial_nplstsq(X, Y, degree, x_pad=10, X_unknown=None):
"""
Fits a polynomial of degree with variables X over Y
:param X: variables. n dimensional array, with row-wise instance
:param Y: value to fit to. n dimensional array, with row wise instances
:param degree: degree of polynomial. list, dim(X) = len(degree)
:param x_pad: padding around X for generated data for the model
:param X_unknown: unknown X for which y is to be predicted, if not None.
:return: W, coefficients.
:return: (x, y): (x, y) values of the generated data for the model.
"""
# TODO: tests for truly n-dimensional X
# construct the appropriate Vandermonde matrix
V = pol.polyvander(X, degree)
# get coefficients by least squares
coeff = la.lstsq(V, Y)[0] # discarding other information
# construct some fitting data
x = np.linspace(X.min() - x_pad, X.max() + x_pad, 100)
x_ = pol.polyvander(x, degree)
y = np.dot(x_, coeff)
# calculate y for unknown X, if provided
if X_unknown is not None:
V_unknown = pol.polyvander(X_unknown, degree)
y_pred = np.dot(V_unknown, coeff)
return coeff, (x, y), (X_unknown, y_pred)
else:
return coeff, (x, y)
def fitting(x, y, deg):
x = np.asarray(x) + 0.0
y = np.asarray(y) + 0.0
deg = np.asarray(deg)
if deg.ndim == 0:
lm = deg
order = lm + 1
van = poly.polyvander(x, lm)
else:
deg = np.sort(deg)
lm = deg[-1]
order = len(deg)
van = poly.polyvander(x, lm)[..., deg]
ls = van.T
rs = y.T
scl = np.sqrt(np.square(ls).sum(1))
scl[scl == 0] = 1
coeff, res, rank, s = la.lstsq(ls.T / scl, rs.T, rcond=len(x)*np.finfo(x.dtype).eps)
coeff = (coeff.T/scl).T
if deg.ndim > 0:
if coeff.ndim == 2:
exp = np.zeros((lm + 1, coeff.shape[1]), dtype=coeff.dtype)
else:
exp = np.zeros(lm + 1, dtype=coeff.dtype)
exp[deg] = coeff
coeff = exp
return coeff
def deriv(self, m, v=None):
alpha = self.slope
sig1, sig2, c = m[0], m[1], m[2:]
if self.logSigma:
sig1, sig2 = np.exp(sig1), np.exp(sig2)
# 2D
if self.mesh.dim == 2:
X = self.mesh.gridCC[self.actInd, 0]
Y = self.mesh.gridCC[self.actInd, 1]
if self.normal == 'X':
f = polynomial.polyval(Y, c) - X
V = polynomial.polyvander(Y, len(c) - 1)
elif self.normal == 'Y':
f = polynomial.polyval(X, c) - Y
V = polynomial.polyvander(X, len(c) - 1)
else:
raise (Exception("Input for normal = X or Y or Z"))
# 3D
elif self.mesh.dim == 3:
X = self.mesh.gridCC[self.actInd, 0]
Y = self.mesh.gridCC[self.actInd, 1]
Z = self.mesh.gridCC[self.actInd, 2]
if self.normal == 'X':
f = (polynomial.polyval2d(
Y, Z, c.reshape(
(self.order[0] + 1, self.order[1] + 1))) - X)
V = polynomial.polyvander2d(Y, Z, self.order)
elif self.normal == 'Y':
f = (polynomial.polyval2d(
X, Z, c.reshape(
(self.order[0] + 1, self.order[1] + 1))) - Y)
V = polynomial.polyvander2d(X, Z, self.order)
elif self.normal == 'Z':
f = (polynomial.polyval2d(
X, Y, c.reshape(
(self.order[0] + 1, self.order[1] + 1))) - Z)
V = polynomial.polyvander2d(X, Y, self.order)
else:
raise (Exception("Input for normal = X or Y or Z"))
if self.logSigma:
g1 = -(np.arctan(alpha * f) / np.pi + 0.5) * sig1 + sig1
g2 = (np.arctan(alpha * f) / np.pi + 0.5) * sig2
else:
g1 = -(np.arctan(alpha * f) / np.pi + 0.5) + 1.0
g2 = (np.arctan(alpha * f) / np.pi + 0.5)
g3 = Utils.sdiag(alpha * (sig2 - sig1) / (1. +
(alpha * f)**2) / np.pi) * V
if v is not None:
return sp.csr_matrix(np.c_[g1, g2, g3]) * v
return sp.csr_matrix(np.c_[g1, g2, g3])
def deriv(self, m, v=None):
alpha = self.slope
sig1, sig2, c = m[0], m[1], m[2:]
if self.logSigma:
sig1, sig2 = np.exp(sig1), np.exp(sig2)
# 2D
if self.mesh.dim == 2:
X = self.mesh.gridCC[self.actInd, 0]
Y = self.mesh.gridCC[self.actInd, 1]
if self.normal == 'X':
f = polynomial.polyval(Y, c) - X
V = polynomial.polyvander(Y, len(c)-1)
elif self.normal == 'Y':
f = polynomial.polyval(X, c) - Y
V = polynomial.polyvander(X, len(c)-1)
else:
raise(Exception("Input for normal = X or Y or Z"))
# 3D
elif self.mesh.dim == 3:
X = self.mesh.gridCC[self.actInd, 0]
Y = self.mesh.gridCC[self.actInd, 1]
Z = self.mesh.gridCC[self.actInd, 2]
if self.normal == 'X':
f = (polynomial.polyval2d(Y, Z, c.reshape((self.order[0]+1,
self.order[1]+1))) - X)
V = polynomial.polyvander2d(Y, Z, self.order)
elif self.normal == 'Y':
f = (polynomial.polyval2d(X, Z, c.reshape((self.order[0]+1,
self.order[1]+1))) - Y)
V = polynomial.polyvander2d(X, Z, self.order)
elif self.normal == 'Z':
f = (polynomial.polyval2d(X, Y, c.reshape((self.order[0]+1,
self.order[1]+1))) - Z)
V = polynomial.polyvander2d(X, Y, self.order)
else:
raise(Exception("Input for normal = X or Y or Z"))
if self.logSigma:
g1 = -(np.arctan(alpha*f)/np.pi + 0.5)*sig1 + sig1
g2 = (np.arctan(alpha*f)/np.pi + 0.5)*sig2
else:
g1 = -(np.arctan(alpha*f)/np.pi + 0.5) + 1.0
g2 = (np.arctan(alpha*f)/np.pi + 0.5)
g3 = Utils.sdiag(alpha*(sig2-sig1)/(1.+(alpha*f)**2)/np.pi)*V
if v is not None:
return sp.csr_matrix(np.c_[g1, g2, g3]) * v
return sp.csr_matrix(np.c_[g1, g2, g3])
def test_polyvander(self) :
# check for 1d x
x = np.arange(3)
v = poly.polyvander(x, 3)
assert_(v.shape == (3,4))
for i in range(4) :
coef = [0]*i + [1]
assert_almost_equal(v[...,i], poly.polyval(x, coef))
# check for 2d x
x = np.array([[1,2],[3,4],[5,6]])
v = poly.polyvander(x, 3)
assert_(v.shape == (3,2,4))
for i in range(4) :
coef = [0]*i + [1]
assert_almost_equal(v[...,i], poly.polyval(x, coef))
def test_polyvander(self) :
# check for 1d x
x = np.arange(3)
v = poly.polyvander(x, 3)
assert_(v.shape == (3,4))
for i in range(4) :
coef = [0]*i + [1]
assert_almost_equal(v[...,i], poly.polyval(x, coef))
# check for 2d x
x = np.array([[1,2],[3,4],[5,6]])
v = poly.polyvander(x, 3)
assert_(v.shape == (3,2,4))
for i in range(4) :
coef = [0]*i + [1]
assert_almost_equal(v[...,i], poly.polyval(x, coef))
def shoot(self, J_c, rho):
'''
Estimate the harmonic frequencies with the passed on a priori.
-----
Arguments:
J_c: int
Number of harmonics in the designated band.
rho: float number or 1D float array
Power spectral density of the noise in the designated band.
-----
Returns:
fhat: float 1D array
Estimate of the harmonic frequencies in the designated band in ascending order.
val: float 1D array
Eigenvalues of the estimate of the harmonic term in descending order.
'''
Z = polyvander(
np.exp(2j * np.pi * self.freq[self.idx_l:self.idx_u + 1]),
J_c + 1).T[1:]
Q = null_space(Z)
D = np.diag(self._N * rho * np.sum(np.abs(Q)**2, axis=1))
YQ = self.y_k[:, self.idx_l:self.idx_u + 1,
np.newaxis] * Q[np.newaxis, :, :]
YPYh = np.mean(YQ @ YQ.conj().transpose(0, 2, 1), axis=0)
val, Vec = eigh(
YPYh - D
) # eigenvalues in ascending order and the corresponding eigenvectors
C_c = Z @ Vec[:, :-J_c - 1:-1]
A_c = self._tls(C_c[:-1], C_c[1:])
fhat = np.angle(eigvals(A_c)) / (2 * np.pi)
return np.sort(fhat), val[::-1]
def fit_polynomial_bayesian_skl(X, Y, degree,
lambda_shape=1.e-6, lambda_invscale=1.e-6,
padding=10, n=100,
X_unknown=None):
X_v = pol.polyvander(X, degree)
clf = BayesianRidge(lambda_1=lambda_shape, lambda_2=lambda_invscale)
clf.fit(X_v, Y)
coeff = np.copy(clf.coef_)
# there some weird intercept thing
# since the Vandermonde matrix has 1 at the beginning, just add this
# intercept to the first coeff
coeff[0] += clf.intercept_
ret_ = [coeff]
# generate the line
x = np.linspace(X.min()-padding, X.max()+padding, n)
x_v = pol.polyvander(x, degree)
# using the provided predict method
y_1 = clf.predict(x_v)
# using np.dot() with coeff
y_2 = np.dot(x_v, coeff)
ret_.append(((x, y_1), (x, y_2)))
if X_unknown is not None:
xu_v = pol.polyvander(X_unknown, degree)
# using the predict method
yu_1 = clf.predict(xu_v)
# using np.dot() with coeff
yu_2 = np.dot(xu_v, coeff)
ret_.append(((X_unknown, yu_1), (X_unknown, yu_2)))
return ret_
def newtVander5d(locations, deg):
# stolen straight from legvander3d and modified
n=locations.shape[1]
ideg = [int(d) for d in deg]
is_valid = [id == d and id >= 0 for id, d in zip(ideg, deg)]
if is_valid != [True,]*n:
raise ValueError("degrees must be non-negative integers")
vi=[poly.polyvander(locations[:,i], deg[i]) for i in range(n)]
v = vi[0][..., None, None, None, None]*vi[1][..., None,:,None, None, None]*vi[2][..., None, None,:,None,None]*vi[3][...,None,None,None,:,None]*vi[4][...,None,None,None,None,:]
return v.reshape(v.shape[:-n] + (-1,))
def genericNewtVander(locations, deg):
# stolen straight from legvander3d and modified
n=locations.shape[1]
ideg = [int(d) for d in deg]
is_valid = [id == d and id >= 0 for id, d in zip(ideg, deg)]
if is_valid != [True,]*n:
raise ValueError("degrees must be non-negative integers")
vi=[poly.polyvander(locations[:,i], deg[i]) for i in range(n)]
indexingTuples=[]
raise NotImplementedError
# v = vx[..., None, None]*vy[..., None,:, None]*vz[..., None, None,:]
return v.reshape(v.shape[:-n] + (-1,))
def error_vandermonde(x, residuals=None, rank=None, *args, **kwargs):
'''
Function to generate 1) error estimation on parameters determined by the function
coef, [residuals, rank, singular_values, rcond] = numpy.polynomial.polynomial.polyfit()
2) covariance matrix associated.
The Vandermonde matrix generated by vand = polyvander(x,rank)
The Covariance matrix cov is obtained via the vandermonde matrix V
via this numerical steps:
1) compute np.dot(V.T,V).inv
2) and multiply it by the residual/(nb of data points - nb of coefficients)
The error parameters are then computed via : error_parameters = np.sqrt(np.diag(cov))
Script written by Timothée Chauviré (https://github.com/TChauvire/EPR_ESR_Suite/), 09/09/2020
Parameters
----------
residuals : first value generated by polyfit in the list of the second output
DESCRIPTION. float number. The default is None.
vandermonde : Vandermonde matrix generated by vand = polyvander(x,rank)
DESCRIPTION. The default is None.
rank : necessary for multidimensional array
DESCRIPTION. The default is None.
Raises
------
ValueError if rank is higher than the number of points
"the number of data points must exceed order "
"to scale the covariance matrix".
Returns
-------
error_parameters : error uncertainties estimated on the parameters ordered
from low to high polynomial order.
By example for a linear model f(x)=a0+a1*x :
error_parameters[0] = constant parameters a0,
error_parameters[1] = slope a1.
cov : covariance estimated for the Least-squares fit of a polynomial to data
return a Matrix(Rank by rank)
'''
if len(x) <= rank:
raise ValueError("the number of data points must exceed order "
"to scale the covariance matrix")
else:
v = polyvander(x, rank) # generate the vandermonde matrix
cov = residuals / (len(x) - rank) * np.linalg.inv(np.dot(v.T, v))
error_parameters = np.sqrt(np.diag(cov))
return error_parameters, cov
def mls2d(x, y, p=0, support_radius=1, deg: int = 1):
"""
двумерный moving least squares
:param x: данные по x
:param y: данные по y соответствующие x
:param p: опорная точка
:param support_radius: радиус влияния
:param deg: степень полинома
:return: коэффициенты полинома
"""
B = polyvander(x, deg)
rj = np.sqrt(np.power(x - p, 2)) / support_radius
W = np.diag(np.where(rj <= 1, 1 - 6 * rj ** 2 + 8 * rj ** 3 - 3 * rj ** 4, 0))
c = inv(B.T @ W @ B) @ B.T @ W @ y
return c
def V(self, Y): """ Build a generalized multivariate Vandermonde matrix for this basis """ assert Y.shape[1] == self.n V_coordinate = [polyvander(Y[:,k], self.p) for k in range(self.n)] V = np.zeros((Y.shape[0], self.N)) for j, terms in enumerate(self.basis_terms): for alpha, coeff in terms: # determine the coefficients on the monomial polynomial # Compute the product of the V_col = np.ones(V.shape[0]) for k in range(0, self.n): V_col *= V_coordinate[k][:,alpha[k]] V[:,j] += coeff * V_col return V
def DV(self, Y): """ Build a generalized multivariate Vandermonde matrix for this basis """ M = Y.shape[0] assert Y.shape[1] == self.n V_coordinate = [polyvander(Y[:,k], self.p) for k in range(self.n)] DV = np.zeros((M, self.N, self.n)) for k in range(self.n): for j, terms in enumerate(self.basis_terms_der[k]): for alpha, coeff in terms: # determine the coefficients on the monomial polynomial # Compute the product of the V_col = np.ones(M) for i in range(self.n): V_col *= V_coordinate[i][:,alpha[i]] DV[:,j,k] += coeff * V_col return DV
def fit_polynomial_bayesian(x, y, degree,
sig2=None, sig2_0=3.0,
use_pinv=False, use_lsmr=False,
padding=10, n=100, get_pdf=True,
X_unknown=None):
X = pol.polyvander(x, degree)
if sig2 is None:
sig2 = np.var(y)
Xt = X.T
prec = (1/sig2) * (np.dot(Xt, X)) + (1/sig2_0) * np.identity(degree + 1)
# different approaches to inverse calculation
# TODO: @motjuste: which one is correct
prec_inv = la.pinv(prec) if use_pinv else la.inv(prec)
mu = (1/sig2) * np.dot(prec_inv, np.dot(Xt, y))
# FIXME: @motjuste: choose the one righteous approach
if use_lsmr:
from scipy.sparse.linalg import lsmr
coeff = lsmr(X, y, damp=(sig2/sig2_0))[0]
else:
if use_pinv:
coeff = la.pinv(np.dot(Xt, X) + (sig2/sig2_0) * np.identity(
degree + 1))
else:
coeff = la.inv(np.dot(Xt, X) + (sig2/sig2_0) * np.identity(
degree + 1))
coeff = np.dot(coeff, np.dot(Xt, y))
ret_ = [coeff]
# generate the line
x = np.linspace(x.min()-padding, x.max()+padding, n)
x_v = pol.polyvander(x, degree)
# the mean of the posterior of y is the best prediction
y_1 = np.dot(x_v, mu)
# using np.dot() with coeff
y_2 = np.dot(x_v, coeff)
ret_.append(((x, y_1), (x, y_2)))
if X_unknown is not None:
xu_v = pol.polyvander(X_unknown, degree)
# the mean of the posterior of y is the best prediction
yu_1 = np.dot(xu_v, mu)
# using np.dot() with coeff
yu_2 = np.dot(xu_v, coeff)
ret_.append(((X_unknown, yu_1), (X_unknown, yu_2)))
if get_pdf:
# http://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.stats.multivariate_normal.html
data_min = np.array([x.min(), y.min()])
data_max = np.array([x.max(), y.max()])
xy_steps = (data_max - data_min) / (n + 2 * padding)
# noinspection PyPep8
xx, yy = np.mgrid[data_min[0] - padding: data_max[0] + padding:
xy_steps[0],
data_min[1] - padding: data_max[1] + padding:
xy_steps[1]]
x = xx[:, 0]
y = yy[0, :]
pdf = []
for i, x_ in enumerate(x):
x_v = pol.polyvander(x_, degree).T
mean = np.dot(mu.T, x_v)
var = sig2 + np.dot(x_v.T, np.dot(prec_inv, x_v))
pdf_ = norm.pdf(y, mean, var).T
pdf.append(pdf_)
pdf = np.array(pdf)[:, :, 0]
# use `plt.contourf(x, y, pdf)`
ret_.append((xx, yy, pdf))
return ret_
def polyvander(x, deg):
from numpy.polynomial.polynomial import polyvander
return polyvander(x, deg)
n = 10
x = np.linspace(xmin, xmax, 100)
# method 1:
# regression using ployfit
c = poly.polyfit(hgt, wgt, n)
y = poly.polyval(x, c)
axs1.set_title("ployfit")
axs1.plot(hgt, wgt, 'ko', x, y, 'r-')
axs1.set_xlim(xmin, xmax)
axs1.set_ylim(ymin, ymax)
# method 2:
# regression using the Vandermonde matrix and pinv
X = poly.polyvander(hgt, n)
c = np.dot(la.pinv(X), wgt)
y = np.dot(poly.polyvander(x, n), c)
axs2.set_title("Vandermonde and pinv")
axs2.plot(hgt, wgt, 'ko', x, y, 'r-')
axs2.set_xlim(xmin, xmax)
axs2.set_ylim(ymin, ymax)
# method 3:
# regression using the Vandermonde matrix and lstsq
X = poly.polyvander(hgt, n)
c = la.lstsq(X, wgt)[0]
y = np.dot(poly.polyvander(x, n), c)
def error_vandermonde(x,residuals=None,rank=None,*args,**kwargs):
'''
Function to generate 1) error estimation on parameters determined by the function
coef, [residuals, rank, singular_values, rcond] = numpy.polynomial.polynomial.polyfit()
2) covariance matrix associated.
The Vandermonde matrix generated by vand = polyvander(x,rank)
The Covariance matrix cov is obtained via the vandermonde matrix V
via this numerical steps:
1) compute np.dot(V.T,V).inv
2) and multiply it by the residual/(nb of data points - nb of coefficients)
The error parameters are then computed via : error_parameters = np.sqrt(np.diag(cov))
Script written by Timothée Chauviré (https://github.com/TChauvire/EPR_ESR_Suite/), 09/09/2020
Parameters
----------
residuals : first value generated by polyfit in the list of the second output
DESCRIPTION. float number. The default is None.
vandermonde : Vandermonde matrix generated by vand = polyvander(x,rank)
DESCRIPTION. The default is None.
rank : necessary for multidimensional array
DESCRIPTION. The default is None.
Raises
------
ValueError if rank is higher than the number of points
"the number of data points must exceed order "
"to scale the covariance matrix".
Returns
-------
error_parameters : error uncertainties estimated on the parameters ordered
from low to high polynomial order.
By example for a linear model f(x)=a0+a1*x :
error_parameters[0] = constant parameters a0,
error_parameters[1] = slope a1.
cov : covariance estimated for the Least-squares fit of a polynomial to data
return a Matrix(Rank by rank)
'''
from numpy.polynomial.polynomial import polyvander
if len(x) <= rank:
raise ValueError("the number of data points must exceed order "
"to scale the covariance matrix")
# note, this used to be: fac = resids / (len(x) - order - 2.0)
# it was deciced that the "- 2" (originally justified by "Bayesian
# uncertainty analysis") is not was the user expects
# (see gh-11196 and gh-11197)
else:
v = polyvander(x,rank) # generate the vandermonde matrix
cov = residuals/(len(x) - rank)*np.linalg.inv(np.dot(v.T, v))
error_parameters=np.sqrt(np.diag(cov))
return error_parameters,cov
# Ideas To Do or test:
# r,p = scipy.stats.pearsonr(x, y) # and other optional statistical test
# reference https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.pearsonr.html#scipy.stats.pearsonr
# https://en.wikipedia.org/wiki/Pearson_correlation_coefficient
# cov = np.cov(x,y)
# Code from scipy.optimize.curve_fit: (https://github.com/scipy/scipy/blob/v1.5.1/scipy/optimize/minpack.py#L532-L834)
# Do Moore-Penrose inverse discarding zero singular values.
# _, s, VT = svd(res.jac, full_matrices=False)
# threshold = np.finfo(float).eps * max(res.jac.shape) * s[0]
# s = s[s > threshold]
# VT = VT[:s.size]
# pcov = np.dot(VT.T / s**2, VT)
# Check lmfit examples :
# https://github.com/lmfit/lmfit-py/blob/master/examples/example_fit_with_derivfunc.py
# bayesian analysis in python
def trsf(x):
return x / 100.
n = 10
x = np.linspace(xmin, xmax, 100)
# method 1:
# regression using ployfit
c = poly.polyfit(hgt, wgt, n)
y = poly.polyval(x, c)
plot_data_and_fit(hgt, wgt, x, y)
# method 2:
# regression using the Vandermonde matrix and pinv
X = poly.polyvander(hgt, n)
c = np.dot(la.pinv(X), wgt)
y = np.dot(poly.polyvander(x, n), c)
plot_data_and_fit(hgt, wgt, x, y)
# method 3:
# regression using the Vandermonde matrix and lstsq
X = poly.polyvander(hgt, n)
c = la.lstsq(X, wgt)[0]
y = np.dot(poly.polyvander(x, n), c)
plot_data_and_fit(hgt, wgt, x, y)
# method 4:
# regression on transformed data using the Vandermonde
# matrix and either pinv or lstsq
X = poly.polyvander(trsf(hgt), n)
def polyvander(x, deg) :
from numpy.polynomial.polynomial import polyvander
return polyvander(x, deg)
def ofit(x, y, deg, rcond=None, full=False, w=None, cov=False): # numpy/polynomial.py
order = int(deg) + 1
x = NX.asarray(x) + 0.0
y = NX.asarray(y) + 0.0
# check arguments.
if deg < 0:
raise ValueError("expected deg >= 0")
if x.ndim != 1:
raise TypeError("expected 1D vector for x")
if x.size == 0:
raise TypeError("expected non-empty vector for x")
if y.ndim < 1 or y.ndim > 2:
raise TypeError("expected 1D or 2D array for y")
if x.shape[0] != y.shape[0]:
raise TypeError("expected x and y to have same length")
# set rcond
if rcond is None:
rcond = len(x)*np.finfo(x.dtype).eps
# set up least squares equation for powers of x
lhs = poly.polyvander(x, order)
rhs = y
# apply weighting
if w is not None:
w = NX.asarray(w) + 0.0
if w.ndim != 1:
raise TypeError("expected a 1-d array for weights")
if w.shape[0] != y.shape[0]:
raise TypeError("expected w and y to have the same length")
lhs *= w[:, NX.newaxis]
if rhs.ndim == 2:
rhs *= w[:, NX.newaxis]
else:
rhs *= w
# scale lhs to improve condition number and solve
scale = NX.sqrt((lhs*lhs).sum(axis=0))
lhs /= scale
c, resids, rank, s = la.lstsq(lhs, rhs, rcond)
c = (c.T/scale).T # broadcast scale coefficients
# warn on rank reduction, which indicates an ill conditioned matrix
if full:
return c, resids, rank, s, rcond
elif cov:
Vbase = inv(dot(lhs.T, lhs))
Vbase /= NX.outer(scale, scale)
if cov == "unscaled":
fac = 1
else:
if len(x) <= order:
raise ValueError("the number of data points must exceed order "
"to scale the covariance matrix")
# note, this used to be: fac = resids / (len(x) - order - 2.0)
# it was deciced that the "- 2" (originally justified by "Bayesian
# uncertainty analysis") is not was the user expects
# (see gh-11196 and gh-11197)
fac = resids / (len(x) - order)
if y.ndim == 1:
return c, Vbase * fac
else:
return c, Vbase[:,:, NX.newaxis] * fac
else:
return c
def savgol_filter_werror(y, window_length, degree, error=None, cov=None,
deriv=None):
ynew = y * 0.0
# Check that window_length is odd
if window_length % 2 == 0:
print("Window length must be odd\n")
exit(11)
# Take care that the window does not spill out of our array
margin = int(window_length/2)
xarr = np.arange(-margin, margin+1)
if cov is not None:
vander = polyvander(xarr, deg=degree)
vanderT = np.transpose(vander)
else:
weight = 1./error
for i in range(margin, y.size-margin):
if cov is None:
z = solve_polyfit(xarr,
y[i-margin:i+margin+1],
degree,
weight[i- margin:i+margin+1],
deriv=deriv)
else:
z = solve_leastsq(y[i-margin:i+margin+1],
cov[i-margin:i+margin+1,i-margin:i+margin+1],
vander,
vanderT,
deriv=deriv)
ynew[i] = P.polyval(0.0, z)
# Now fit the left boundary, by fitting the first window_length points with
# a degree order polynomial
if cov is None:
z = solve_polyfit(xarr,
y[:window_length],
degree,
weight[:window_length],
deriv=deriv)
else:
z = solve_leastsq(y[:window_length],
cov[:window_length, :window_length],
vander,
vanderT,
deriv=deriv)
for i in range(margin):
ynew[i] = P.polyval(xarr[i], z)
# Now fit the right boundary, by fitting the last window_length points with
# a degree order polynomial
if cov is None:
z = solve_polyfit(xarr,
y[-window_length:],
degree,
weight[-window_length:],
deriv=deriv)
else:
z = solve_leastsq(y[-window_length:],
cov[-window_length:, -window_length:],
vander,
vanderT,
deriv=deriv)
for i in range(margin):
ynew[y.size-margin+i] = P.polyval(xarr[i+margin+1], z)
return ynew
x = np.linspace(xmin, xmax, 10000)
# ---------
# method 1:
# regression using ployfit
c = poly.polyfit(hgt, wgt, n)
y = poly.polyval(x, c)
plot_data_and_fit(hgt, wgt, x, y)
print("polyfit error = ", get_error(hgt, wgt, x, y))
# ---------
# method 2:
# regression using the Vandermonde matrix and pinv
X = poly.polyvander(hgt, n)
c = np.dot(la.pinv(X), wgt)
# print("vander condition: ",la.cond(X))
# print("pinv condition: ",la.cond(la.pinv(X)))
y = np.dot(poly.polyvander(x, n), c)
plot_data_and_fit(hgt, wgt, x, y)
print("vander+pinv error = ", get_error(hgt, wgt, x, y), la.cond(X))
# ---------
# method 3:
# regression using the Vandermonde matrix and lstsq
X = poly.polyvander(hgt, n)
c = la.lstsq(X, wgt)[0]
y = np.dot(poly.polyvander(x, n), c)
plot_data_and_fit(hgt, wgt, x, y)