def test_nextpow2(self):
"""Test the nextpow2 function."""
self.assertEqual(utils.nextpow2(2), 1)
self.assertEqual(utils.nextpow2(17), 5)
import warnings
with warnings.catch_warnings(record=True) as catcher:
utils.nextpow2(-3)
self.assertEqual(len(catcher), 1)
self.assertTrue(catcher[-1].category, RuntimeWarning)
def test_nextpow2(self):
"""Test the nextpow2 function."""
self.assertEqual(utils.nextpow2(2), 1)
self.assertEqual(utils.nextpow2(17), 5)
import warnings
with warnings.catch_warnings(record=True) as catcher:
utils.nextpow2(-3)
self.assertEqual(len(catcher), 1)
self.assertTrue(catcher[-1].category, RuntimeWarning)
def noisecu(n_points):
"""Compute analytic complex uniform white noise.
:param n_points: Length of the noise signal.
:type n_points: int
:return: analytic complex uniform white noise signal of length N
:rtype: numpy.ndarray
:Examples:
>>> import numpy as np
>>> noise = noisecu(512)
>>> print("%.2f" % abs((noise ** 2).mean()))
0.00
>>> print("%.1f" % np.std(noise) ** 2)
1.0
>>> subplot(211), plot(real(noise)) #doctest: +SKIP
>>> subplot(212), plot(linspace(-0.5, 0.5, 512), abs(fftshift(fft(noise))) ** 2) #doctest: +SKIP
.. plot:: docstring_plots/generators/noise/noisecu.py
"""
if n_points <= 2:
noise = (np.random.rand(n_points, 1) - 0.5 + 1j * (np.random.rand(n_points, 1) - 0.5)) * np.sqrt(6)
else:
noise = np.random.rand(2 ** int(nextpow2(n_points)),) - 0.5
noise = hilbert(noise) / noise.std() / np.sqrt(2)
inds = noise.shape[0] - np.arange(n_points - 1, -1, step=-1) - 1
noise = noise[inds]
return noise
def noisecg(n_points, a1=None, a2=None):
"""
Generate analytic complex gaussian noise with mean 0.0 and variance 1.0.
:param n_points: Length of the desired output signal.
:param a1:
Coefficients of the filter through which the noise is passed.
:param a2:
Coefficients of the filter through which the noise is passed.
:type n_points: int
:type a1: float
:type a2: float
:return: Analytic complex Gaussian noise of length n_points.
:rtype: numpy.ndarray
:Examples:
>>> import numpy as np
>>> noise = noisecg(512)
>>> print("%.2f" % abs((noise ** 2).mean()))
0.00
>>> print("%.1f" % np.std(noise) ** 2)
1.0
>>> subplot(211), plot(real(noise)) #doctest: +SKIP
>>> subplot(212), plot(linspace(-0.5, 0.5, 512), abs(fftshift(fft(noise))) ** 2) #doctest: +SKIP
.. plot:: docstring_plots/generators/noise/noisecg.py
"""
assert n_points > 0
if n_points <= 2:
noise = (np.random.randn(n_points, 1.) + 1j * np.random.randn(n_points, 1.)) / np.sqrt(2.)
else:
noise = np.random.normal(size=int(2. ** nextpow2(float(n_points)),))
noise = hilbert(noise) / noise.std() / np.sqrt(2.)
noise = noise[len(noise) - np.arange(n_points - 1, -1, -1) - 1]
return noise
def noisecu(n_points):
"""Compute analytic complex uniform white noise.
:param n_points: Length of the noise signal.
:type n_points: int
:return: analytic complex uniform white noise signal of length N
:rtype: numpy.ndarray
:Examples:
>>> noise = noisecu(512)
>>> print noise.mean()
0.0
>>> print std(noise) ** 2
1.0
>>> subplot(211), plot(real(noise))
>>> subplot(212), plot(linspace(-0.5, 0.5, 512), abs(fftshift(fft(noise))) ** 2)
.. plot:: docstring_plots/generators/noise/noisecu.py
"""
if n_points <= 2:
noise = (np.random.rand(n_points, 1) - 0.5 + 1j *
(np.random.rand(n_points, 1) - 0.5)) * np.sqrt(6)
else:
noise = np.random.rand(2**nextpow2(n_points), ) - 0.5
noise = hilbert(noise) / noise.std() / np.sqrt(2)
inds = noise.shape[0] - np.arange(n_points - 1, -1, step=-1) - 1
noise = noise[inds]
return noise
def noisecg(n_points, a1=None, a2=None):
"""
Generate analytic complex gaussian noise with mean 0.0 and variance 1.0.
:param n_points: Length of the desired output signal.
:param a1:
Coefficients of the filter through which the noise is passed.
:param a2:
Coefficients of the filter through which the noise is passed.
:type n_points: int
:type a1: float
:type a2: float
:return: Analytic complex Gaussian noise of length n_points.
:rtype: numpy.ndarray
:Examples:
>>> noise = noisecg(512)
>>> print noise.mean()
0.0
>>> print std(noise) ** 2
1.0
>>> subplot(211), plot(real(noise))
>>> subplot(212), plot(linspace(-0.5, 0.5, 512), abs(fftshift(fft(noise))) ** 2)
.. plot:: docstring_plots/generators/noise/noisecg.py
"""
assert n_points > 0
if n_points <= 2:
noise = (np.random.randn(n_points, 1) +
1j * np.random.randn(n_points, 1)) / np.sqrt(2)
else:
noise = np.random.normal(size=(2**nextpow2(n_points), ))
noise = hilbert(noise) / noise.std() / np.sqrt(2)
noise = noise[len(noise) - np.arange(n_points - 1, -1, -1) - 1]
return noise
def noisecu(n_points):
"""Compute analytic complex uniform white noise.
:param n_points: Length of the noise signal.
:type n_points: int
:return: analytic complex uniform white noise signal of length N
:rtype: numpy.ndarray
:Examples:
>>> import matplotlib.pyplot as plt
>>> import numpy as np
>>> noise = noisecu(512)
>>> print("%.2f" % abs((noise ** 2).mean()))
0.00
>>> print("%.1f" % np.std(noise) ** 2)
1.0
>>> plt.subplot(211), plt.plot(real(noise)) #doctest: +SKIP
>>> plt.subplot(212), #doctest: +SKIP
>>> plt.plot(linspace(-0.5, 0.5, 512), abs(fftshift(fft(noise))) ** 2) #doctest: +SKIP
.. plot:: docstring_plots/generators/noise/noisecu.py
"""
if n_points <= 2:
noise = (np.random.rand(n_points, 1) - 0.5 + 1j * (np.random.rand(n_points, 1) - 0.5)) * \
np.sqrt(6)
else:
noise = np.random.rand(2**int(nextpow2(n_points)), ) - 0.5
noise = hilbert(noise) / noise.std() / np.sqrt(2)
inds = noise.shape[0] - np.arange(n_points - 1, -1, step=-1) - 1
noise = noise[inds]
return noise
def _get_nvoices(self):
q = (self.bw * self.T * (1 + 2 / self.R) * np.log(
(1 + self.R / 2) / (1 - self.R / 2)))
nq = np.ceil(q / 2)
nmin = nq - nq % 2
ndflt = 2**nextpow2(nmin)
self.n_voices = int(ndflt)
def wide_band(signal, fmin=None, fmax=None, N=None):
if 1 in signal.shape:
signal = signal.ravel()
elif signal.ndim != 1:
raise ValueError("The input signal should be one dimensional.")
s_ana = hilbert(np.real(signal))
nx = signal.shape[0]
m = np.round(nx / 2.0)
t = np.arange(nx) - m
tmin = 0
tmax = nx - 1
T = tmax - tmin
# determine default values for fmin, fmax
if (fmin is None) or (fmax is None):
from matplotlib.mlab import find
STF = np.fft.fftshift(s_ana)
sp = np.abs(STF[:m]) ** 2
maxsp = np.amax(sp)
f = np.linspace(0, 0.5, m + 1)
f = f[:m]
indmin = find(sp > maxsp / 100.0).min()
indmax = find(sp > maxsp / 100.0).max()
if fmin is None:
fmin = max([0.01, 0.05 * np.fix(f[indmin] / 0.05)])
if fmax is None:
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
B = fmax - fmin
R = B / ((fmin + fmax) / 2.0)
nq = np.ceil((B * T * (1 + 2.0 / R) * np.log((1 + R / 2.0) / (1 - R / 2.0))) / 2.0)
nmin = nq - (nq % 2)
if N is None:
N = int(2 ** (nextpow2(nmin)))
# geometric sampling for the analyzed spectrum
k = np.arange(1, N + 1)
q = (fmax / fmin) ** (1.0 / (N - 1))
geo_f = fmin * (np.exp((k - 1) * np.log(q)))
tfmatx = -2j * np.dot(t.reshape(-1, 1), geo_f.reshape(1, -1)) * np.pi
tfmatx = np.exp(tfmatx)
S = np.dot(s_ana.reshape(1, -1), tfmatx)
S = np.tile(S, (nx, 1))
Sb = S * tfmatx
tau = t
S = np.c_[S, np.zeros((nx, N))].T
Sb = np.c_[Sb, np.zeros((nx, N))].T
# mellin transform computation of the analyzed signal
p = np.arange(2 * N)
coef = np.exp(p / 2.0 * np.log(q))
mellinS = np.fft.fftshift(np.fft.ifft(S[:, 0] * coef))
mellinS = np.tile(mellinS, (nx, 1)).T
mellinSb = np.zeros((2 * N, nx), dtype=complex)
for i in range(nx):
mellinSb[:, i] = np.fft.fftshift(np.fft.ifft(Sb[:, i] * coef))
k = np.arange(1, 2 * N + 1)
scale = np.logspace(np.log10(fmin / fmax), np.log10(fmax / fmin), N)
theta = np.log(scale)
mellinSSb = mellinS * np.conj(mellinSb)
waf = np.fft.ifft(mellinSSb, N, axis=0)
no2 = int((N + N % 2) / 2.0)
waf = np.r_[waf[no2:(N + 1), :], waf[:no2, :]]
# normalization
s = np.real(s_ana)
SP = np.fft.fft(hilbert(s))
indmin = int(1 + np.round(fmin * (nx - 2)))
indmax = int(1 + np.round(fmax * (nx - 2)))
sp_ana = SP[(indmin - 1):indmax]
waf *= (np.linalg.norm(sp_ana) ** 2) / waf[no2 - 1, m - 1] / N
return waf, tau, theta
def bertrand(signal, timestamps=None, fmin=None, fmax=None, n_voices=None):
"""bertrand
:param signal:
:param timestamps:
:param fmin:
:param fmax:
:param n_voices:
:type signal:
:type timestamps:
:type fmin:
:type fmax:
:type n_voices:
:return:
:rtype:
"""
xrow = signal.shape[0]
if timestamps is None:
timestamps = np.arange(xrow)
tcol = timestamps.shape[0]
x1 = signal.copy()
x2 = signal.copy()
s1 = np.real(x1)
s2 = np.real(x2)
m = (xrow + (xrow % 2)) / 2
t = np.arange(xrow) - m - 1
tmin = 1
tmax = xrow
T = tmax - tmin
mt = xrow
if (fmin is None) or (fmax is None):
stf1 = np.fft.fft(
np.fft.fftshift(s1[timestamps.min():timestamps.max() + 1]))
stf2 = np.fft.fft(
np.fft.fftshift(s2[timestamps.min():timestamps.max() + 1]))
nstf = stf1.shape[0]
sp1 = np.abs(stf1[:int(np.round(nstf / 2.0))])**2
sp2 = np.abs(stf2[:int(np.round(nstf / 2.0))])**2
maxsp1 = sp1.max()
maxsp2 = sp2.max()
f = np.linspace(0, 0.5, np.round(nstf / 2.0) + 1)
if fmin is None:
mask = sp1 > maxsp1 / 100.0
indmin = np.arange(mask.shape[0],
dtype=int)[mask.astype(bool)].min()
fmin = max([0.01, 0.05 * np.floor(f[indmin] / 0.05)])
if fmax is None:
mask = sp2 > maxsp2 / 100.0
indmax = np.arange(mask.shape[0],
dtype=int)[mask.astype(bool)].max()
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
bw = fmax - fmin
R = bw / (fmin + fmax) * 2.0
umaxbert = lambda x: np.exp(x) - fmax / fmin
umax = brenth(umaxbert, 0, 4)
teq = m / (fmax * umax)
if teq < mt:
m0 = np.round((2 * m**2) / teq - m) + 1
m1 = m + m0
T = 2 * m1 - 1
else:
m0 = 0
m1 = m
if n_voices is None:
nq = np.ceil((bw * T * (1 + 2.0 / R) * np.log(
(1 + R / 2.0) / (1 - R / 2.0))) / 2)
nmin = nq - nq % 2
ndflt = 2**nextpow2(nmin)
n_voices = int(ndflt)
# Geometric sampling for the analyzed spectrum
k = np.arange(1, n_voices + 1)
q = (fmax / fmin)**(1 / (n_voices - 1.0))
t = np.arange(1, mt + 1) - m - 1
geo_f = fmin * np.exp((k - 1) * np.log(q))
tfmatx = np.exp(
-2 * 1j *
np.dot(t.reshape(t.shape[0], 1), geo_f.reshape(1, geo_f.shape[0])) *
np.pi)
S1 = np.dot(s1.reshape(1, s1.shape[0]), tfmatx)
S2 = np.dot(s2.reshape(1, s2.shape[0]), tfmatx)
S1 = np.append(S1, np.zeros((n_voices, )))
S2 = np.append(S2, np.zeros((n_voices, )))
# Mellin tranform of signal
p = np.arange(2 * n_voices)
mellin1 = np.fft.fftshift(np.fft.ifft(S1))
mellin2 = np.fft.fftshift(np.fft.ifft(S2))
umin = -umax
du = np.abs(umax - umin) / (2 * m1)
u = np.linspace(umin, umax - du, (umax - umin) / du)
u[m1] = 0
beta = (p / float(n_voices) - 1) / (2 * np.log(q))
# Computation of P0(t. f, f)
waf = np.zeros((2 * m1, n_voices), dtype=complex)
for n in np.hstack((np.arange(1, m1 + 1), np.arange(m1 + 2, 2 * m1 + 1))):
mx1 = np.exp((-2 * 1j * np.pi * beta + 0.5) * np.log(
(u[n - 1] / 2) * np.exp(-u[n - 1] / 2.0) /
np.sinh(u[n - 1] / 2))) * mellin1
mx2 = np.exp((-2 * 1j * np.pi * beta + 0.5) * np.log(
(u[n - 1] / 2) * np.exp(u[n - 1] / 2.0) /
np.sinh(u[n - 1] / 2))) * mellin2
fx1 = np.fft.fft(np.fft.fftshift(mx1))[:n_voices]
fx2 = np.fft.fft(np.fft.fftshift(mx2))[:n_voices]
waf[n - 1, :] = fx1 * np.conj(fx2)
waf[m1, :] = S1[:n_voices] * np.conj(S2[:n_voices])
waf = np.vstack((waf[m1:(2 * m1), :], waf[:m1, :]))
waf *= np.repeat(geo_f.reshape((1, geo_f.shape[0])), 2 * m1, axis=0)
tffr = np.fft.ifft(waf, axis=0)
tffr = np.real(
np.rot90(np.vstack((tffr[m1:(2 * m1 + 1), :], tffr[:m1, :])), k=-1))
# conversion from tff to tf using 1d interpolation
tfr = np.zeros((n_voices, tcol))
ts2 = (mt - 1.0) / 2
gamma = np.linspace(-geo_f[n_voices - 1] * ts2, geo_f[n_voices - 1] * ts2,
2 * m1)
for i in xrange(n_voices):
ind = find(
np.logical_and(gamma >= -geo_f[i] * ts2, gamma <= geo_f[i] * ts2))
x = gamma[ind]
y = tffr[i, ind]
xi = (timestamps - ts2 - 1) * geo_f[i]
tck = splrep(x, y)
tfr[i, :] = splev(xi, tck).ravel()
t = timestamps
f = geo_f.ravel()
# Normalization
SP1 = np.fft.fft(hilbert(s1), axis=0)
SP2 = np.fft.fft(hilbert(s2), axis=0)
indmin = 1 + int(np.round(fmin * (tcol - 2)))
indmax = 1 + int(np.round(fmax * (tcol - 2)))
sp1_ana = SP1[(indmin - 1):indmax]
sp2_ana = SP2[(indmin - 1):indmax]
tfr = tfr * np.dot(sp1_ana.T, sp2_ana) / integrate_2d(tfr, t, f) / n_voices
return tfr, t, f
def smoothed_pseudo_wigner(signal,
timestamps=None,
K='bertrand',
nh0=None,
ng0=0,
fmin=None,
fmax=None,
n_voices=None):
"""smoothed_pseudo_wigner
:param signal:
:param timestamps:
:param K:
:param nh0:
:param ng0:
:param fmin:
:param fmax:
:param n_voices:
:type signal:
:type timestamps:
:type K:
:type nh0:
:type ng0:
:type fmin:
:type fmax:
:type n_voices:
:return:
:rtype:
"""
xrow = signal.shape[0]
if timestamps is None:
timestamps = np.arange(signal.shape[0])
if nh0 is None:
nh0 = np.sqrt(signal.shape[0])
tcol = timestamps.shape[0]
mt = signal.shape[0]
x1 = x2 = signal.copy()
s1 = np.real(x1)
s2 = np.real(x2)
m = (mt + np.remainder(mt, 2.0)) / 2.0
if (fmin is None) or (fmax is None):
stf1 = np.fft.fft(
np.fft.fftshift(s1[timestamps.min():timestamps.max() + 1]))
stf2 = np.fft.fft(
np.fft.fftshift(s2[timestamps.min():timestamps.max() + 1]))
nstf = stf1.shape[0]
sp1 = np.abs(stf1[:int(np.round(nstf / 2.0))])**2
sp2 = np.abs(stf2[:int(np.round(nstf / 2.0))])**2
maxsp1 = sp1.max()
maxsp2 = sp2.max()
f = np.linspace(0, 0.5,
np.round(nstf / 2.0) + 1)[:int(np.round(nstf / 2.0))]
if fmin is None:
mask = sp1 > maxsp1 / 100.0
indmin = np.arange(mask.shape[0],
dtype=int)[mask.astype(bool)].min()
fmin = max([0.01, 0.05 * np.floor(f[indmin] / 0.05)])
if fmax is None:
mask = sp2 > maxsp2 / 100.0
indmax = np.arange(mask.shape[0],
dtype=int)[mask.astype(bool)].max()
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
B = fmax - fmin
R = B / ((fmin + fmax) / 2.0)
ratio = fmax / fmin
umax = np.log(ratio)
teq = nh0 / (fmax * umax)
if teq > 2 * nh0:
m0 = (2 * nh0**2) / teq - nh0 + 1
else:
m0 = 0
mu = np.round(nh0 + m0)
T = 2 * mu - 1
if n_voices is None:
nq = np.ceil((B * T * (1 + 2.0 / R) * np.log(
(1 + R / 2.0) / (1 - R / 2.0))) / 2)
nmin = nq - nq % 2
ndflt = 2**nextpow2(nmin)
n_voices = int(ndflt)
k = np.arange(1, n_voices + 1)
q = ratio**(1.0 / (n_voices - 1))
a = np.exp((k - 1) * np.log(q))
geo_f = fmin * a
# Wavelet decomposition computation
matxte1 = np.zeros((n_voices, tcol), dtype=complex)
matxte2 = np.zeros((n_voices, tcol), dtype=complex)
_, _, _, wt1 = scalogram(s1,
time_instants=timestamps,
waveparams=nh0,
fmin=fmin,
fmax=fmax,
n_voices=n_voices)
_, _, _, wt2 = scalogram(s2,
time_instants=timestamps,
waveparams=nh0,
fmin=fmin,
fmax=fmax,
n_voices=n_voices)
for ptr in xrange(n_voices):
matxte1[ptr, :] = wt1[ptr, :] * np.sqrt(a[n_voices - ptr - 1])
matxte2[ptr, :] = wt2[ptr, :] * np.sqrt(a[n_voices - ptr - 1])
umin = -umax
u = np.linspace(umin, umax, 2 * mu + 1)
u = u[:(2 * mu)]
u[mu] = 0
p = np.arange(2 * n_voices)
beta = (p / float(n_voices) - 1.0) / (2 * np.log(q))
l1 = l2 = np.zeros((2 * mu, 2 * n_voices), dtype=complex)
for m in xrange(l1.shape[0]):
l1[m, :] = np.exp(-2 * np.pi * 1j * beta * np.log(lambdak(u[m], K)))
l2[m, :] = np.exp(-2 * np.pi * 1j * beta * np.log(lambdak(-u[m], K)))
# Calculate time smoothing window
if ng0 == 0:
G = np.ones((2 * mu))
else:
a_t = 3
sigma_t = ng0 * fmax / np.sqrt(2 * a_t * np.log(10))
a_u = 2 * np.pi**2 * sigma_t**2 * umax**2 / np.log(10)
G = np.exp(-(a_u * np.log(10) / mu**2) * np.arange(-mu, mu)**2)
waf = np.zeros((2 * mu, n_voices))
tfr = np.zeros((n_voices, tcol))
S1 = S2 = np.zeros((2 * n_voices, ), dtype=complex)
mx1 = mx2 = np.zeros((2 * n_voices, 2 * mu))
for ti in xrange(tcol):
S1[:n_voices] = matxte1[:, ti]
mellin1 = np.fft.fftshift(np.fft.ifft(S1))
mx1 = l1 * mellin1.reshape(1, mellin1.shape[0]).repeat(2 * mu, 0)
mx1 = np.fft.fft(mx1, axis=0)
tx1 = mx1[:n_voices, :].T
S2[:n_voices] = matxte2[:, ti]
mellin2 = np.fft.fftshift(np.fft.ifft(S2))
mx2 = l2 * mellin2.reshape(1, mellin2.shape[0]).repeat(2 * mu, 0)
mx2 = np.fft.fft(mx2, axis=0)
tx2 = mx2[:n_voices, :].T
waf = np.real(tx1 * np.conj(tx2)) * G.reshape(G.shape[0], 1).repeat(
n_voices, axis=1)
tfr[:, ti] = np.sum(waf) * geo_f
t = timestamps
f = geo_f
# Normalization
sp1 = np.fft.fft(hilbert(s1))
sp2 = np.fft.fft(hilbert(s2))
indmin = 1 + np.round(fmin * (xrow - 2))
indmax = 1 + np.round(fmax * (xrow - 2))
sp1_ana = sp1[indmin:(indmax + 1)]
sp2_ana = sp2[indmin:(indmax + 1)]
xx = np.dot(np.real(sp1_ana), np.real(sp2_ana))
xx += np.dot(np.imag(sp1_ana), np.imag(sp2_ana))
tfr = tfr * xx / integrate_2d(tfr, t, f) / n_voices
return tfr, t, f
def bertrand(signal, timestamps=None, fmin=None, fmax=None, n_voices=None):
"""bertrand
:param signal:
:param timestamps:
:param fmin:
:param fmax:
:param n_voices:
:type signal:
:type timestamps:
:type fmin:
:type fmax:
:type n_voices:
:return:
:rtype:
"""
xrow = signal.shape[0]
if timestamps is None:
timestamps = np.arange(xrow)
tcol = timestamps.shape[0]
x1 = signal.copy()
x2 = signal.copy()
s1 = np.real(x1)
s2 = np.real(x2)
m = (xrow + (xrow % 2)) / 2
t = np.arange(xrow) - m - 1
tmin = 1
tmax = xrow
T = tmax - tmin
mt = xrow
if (fmin is None) or (fmax is None):
stf1 = np.fft.fft(np.fft.fftshift(s1[timestamps.min():timestamps.max() + 1]))
stf2 = np.fft.fft(np.fft.fftshift(s2[timestamps.min():timestamps.max() + 1]))
nstf = stf1.shape[0]
sp1 = np.abs(stf1[:int(np.round(nstf / 2.0))]) ** 2
sp2 = np.abs(stf2[:int(np.round(nstf / 2.0))]) ** 2
maxsp1 = sp1.max()
maxsp2 = sp2.max()
f = np.linspace(0, 0.5, np.round(nstf / 2.0) + 1)
if fmin is None:
mask = sp1 > maxsp1 / 100.0
indmin = np.arange(mask.shape[0], dtype=int)[mask.astype(bool)].min()
fmin = max([0.01, 0.05 * np.floor(f[indmin] / 0.05)])
if fmax is None:
mask = sp2 > maxsp2 / 100.0
indmax = np.arange(mask.shape[0], dtype=int)[mask.astype(bool)].max()
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
bw = fmax - fmin
R = bw / (fmin + fmax) * 2.0
umaxbert = lambda x: np.exp(x) - fmax / fmin
umax = brenth(umaxbert, 0, 4)
teq = m / (fmax * umax)
if teq < mt:
m0 = np.round((2 * m ** 2) / teq - m) + 1
m1 = m + m0
T = 2 * m1 - 1
else:
m0 = 0
m1 = m
if n_voices is None:
nq = np.ceil((bw * T * (1 + 2.0 / R) * np.log((1 + R / 2.0) / (1 - R / 2.0))) / 2)
nmin = nq - nq % 2
ndflt = 2 ** nextpow2(nmin)
n_voices = int(ndflt)
# Geometric sampling for the analyzed spectrum
k = np.arange(1, n_voices + 1)
q = (fmax / fmin) ** (1 / (n_voices - 1.0))
t = np.arange(1, mt + 1) - m - 1
geo_f = fmin * np.exp((k - 1) * np.log(q))
tfmatx = np.exp(-2 * 1j * np.dot(t.reshape(t.shape[0], 1),
geo_f.reshape(1, geo_f.shape[0])) * np.pi)
S1 = np.dot(s1.reshape(1, s1.shape[0]), tfmatx)
S2 = np.dot(s2.reshape(1, s2.shape[0]), tfmatx)
S1 = np.append(S1, np.zeros((n_voices,)))
S2 = np.append(S2, np.zeros((n_voices,)))
# Mellin tranform of signal
p = np.arange(2 * n_voices)
mellin1 = np.fft.fftshift(np.fft.ifft(S1))
mellin2 = np.fft.fftshift(np.fft.ifft(S2))
umin = -umax
du = np.abs(umax - umin) / (2 * m1)
u = np.linspace(umin, umax - du, (umax - umin) / du)
u[m1] = 0
beta = (p / float(n_voices) - 1) / (2 * np.log(q))
# Computation of P0(t. f, f)
waf = np.zeros((2 * m1, n_voices), dtype=complex)
for n in np.hstack((np.arange(1, m1 + 1), np.arange(m1 + 2, 2 * m1 + 1))):
mx1 = np.exp((-2 * 1j * np.pi * beta + 0.5) * np.log((u[n - 1] / 2) *
np.exp(-u[n - 1] / 2.0) / np.sinh(u[n - 1] / 2))) * mellin1
mx2 = np.exp((-2 * 1j * np.pi * beta + 0.5) * np.log((u[n - 1] / 2) *
np.exp(u[n - 1] / 2.0) / np.sinh(u[n - 1] / 2))) * mellin2
fx1 = np.fft.fft(np.fft.fftshift(mx1))[:n_voices]
fx2 = np.fft.fft(np.fft.fftshift(mx2))[:n_voices]
waf[n - 1, :] = fx1 * np.conj(fx2)
waf[m1, :] = S1[:n_voices] * np.conj(S2[:n_voices])
waf = np.vstack((waf[m1:(2 * m1), :], waf[:m1, :]))
waf *= np.repeat(geo_f.reshape((1, geo_f.shape[0])), 2 * m1, axis=0)
tffr = np.fft.ifft(waf, axis=0)
tffr = np.real(np.rot90(np.vstack((tffr[m1:(2 * m1 + 1), :],
tffr[:m1, :])), k=-1))
# conversion from tff to tf using 1d interpolation
tfr = np.zeros((n_voices, tcol))
ts2 = (mt - 1.0) / 2
gamma = np.linspace(-geo_f[n_voices - 1] * ts2,
geo_f[n_voices - 1] * ts2, 2 * m1)
for i in range(n_voices):
ind = find(np.logical_and(gamma >= -geo_f[i] * ts2,
gamma <= geo_f[i] * ts2))
x = gamma[ind]
y = tffr[i, ind]
xi = (timestamps - ts2 - 1) * geo_f[i]
tck = splrep(x, y)
tfr[i, :] = splev(xi, tck).ravel()
t = timestamps
f = geo_f.ravel()
# Normalization
SP1 = np.fft.fft(hilbert(s1), axis=0)
SP2 = np.fft.fft(hilbert(s2), axis=0)
indmin = 1 + int(np.round(fmin * (tcol - 2)))
indmax = 1 + int(np.round(fmax * (tcol - 2)))
sp1_ana = SP1[(indmin - 1):indmax]
sp2_ana = SP2[(indmin - 1):indmax]
tfr = tfr * np.dot(sp1_ana.T, sp2_ana) / integrate_2d(tfr, t, f) / n_voices
return tfr, t, f
def smoothed_pseudo_wigner(signal, timestamps=None, K='bertrand', nh0=None,
ng0=0, fmin=None, fmax=None, n_voices=None):
"""smoothed_pseudo_wigner
:param signal:
:param timestamps:
:param K:
:param nh0:
:param ng0:
:param fmin:
:param fmax:
:param n_voices:
:type signal:
:type timestamps:
:type K:
:type nh0:
:type ng0:
:type fmin:
:type fmax:
:type n_voices:
:return:
:rtype:
"""
xrow = signal.shape[0]
if timestamps is None:
timestamps = np.arange(signal.shape[0])
if nh0 is None:
nh0 = np.sqrt(signal.shape[0])
tcol = timestamps.shape[0]
mt = signal.shape[0]
x1 = x2 = signal.copy()
s1 = np.real(x1)
s2 = np.real(x2)
m = (mt + np.remainder(mt, 2.0)) / 2.0
if (fmin is None) or (fmax is None):
stf1 = np.fft.fft(np.fft.fftshift(s1[timestamps.min():timestamps.max() + 1]))
stf2 = np.fft.fft(np.fft.fftshift(s2[timestamps.min():timestamps.max() + 1]))
nstf = stf1.shape[0]
sp1 = np.abs(stf1[:int(np.round(nstf / 2.0))]) ** 2
sp2 = np.abs(stf2[:int(np.round(nstf / 2.0))]) ** 2
maxsp1 = sp1.max()
maxsp2 = sp2.max()
f = np.linspace(0, 0.5, np.round(nstf / 2.0) + 1)[:int(np.round(nstf / 2.0))]
if fmin is None:
mask = sp1 > maxsp1 / 100.0
indmin = np.arange(mask.shape[0], dtype=int)[mask.astype(bool)].min()
fmin = max([0.01, 0.05 * np.floor(f[indmin] / 0.05)])
if fmax is None:
mask = sp2 > maxsp2 / 100.0
indmax = np.arange(mask.shape[0], dtype=int)[mask.astype(bool)].max()
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
B = fmax - fmin
R = B / ((fmin + fmax) / 2.0)
ratio = fmax / fmin
umax = np.log(ratio)
teq = nh0 / (fmax * umax)
if teq > 2 * nh0:
m0 = (2 * nh0 ** 2) / teq - nh0 + 1
else:
m0 = 0
mu = np.round(nh0 + m0)
T = 2 * mu - 1
if n_voices is None:
nq = np.ceil((B * T * (1 + 2.0 / R) * np.log((1 + R / 2.0) / (1 - R / 2.0))) / 2)
nmin = nq - nq % 2
ndflt = 2 ** nextpow2(nmin)
n_voices = int(ndflt)
k = np.arange(1, n_voices + 1)
q = ratio ** (1.0 / (n_voices - 1))
a = np.exp((k - 1) * np.log(q))
geo_f = fmin * a
# Wavelet decomposition computation
matxte1 = np.zeros((n_voices, tcol), dtype=complex)
matxte2 = np.zeros((n_voices, tcol), dtype=complex)
_, _, _, wt1 = scalogram(s1, time_instants=timestamps, waveparams=nh0,
fmin=fmin, fmax=fmax, n_voices=n_voices)
_, _, _, wt2 = scalogram(s2, time_instants=timestamps, waveparams=nh0,
fmin=fmin, fmax=fmax, n_voices=n_voices)
for ptr in range(n_voices):
matxte1[ptr, :] = wt1[ptr, :] * np.sqrt(a[n_voices - ptr - 1])
matxte2[ptr, :] = wt2[ptr, :] * np.sqrt(a[n_voices - ptr - 1])
umin = -umax
u = np.linspace(umin, umax, 2 * mu + 1)
u = u[:(2 * mu)]
u[mu] = 0
p = np.arange(2 * n_voices)
beta = (p / float(n_voices) - 1.0) / (2 * np.log(q))
l1 = l2 = np.zeros((2 * mu, 2 * n_voices), dtype=complex)
for m in range(l1.shape[0]):
l1[m, :] = np.exp(-2 * np.pi * 1j * beta * np.log(lambdak(u[m], K)))
l2[m, :] = np.exp(-2 * np.pi * 1j * beta * np.log(lambdak(-u[m], K)))
# Calculate time smoothing window
if ng0 == 0:
G = np.ones((2 * mu))
else:
a_t = 3
sigma_t = ng0 * fmax / np.sqrt(2 * a_t * np.log(10))
a_u = 2 * np.pi ** 2 * sigma_t ** 2 * umax ** 2 / np.log(10)
G = np.exp(-(a_u * np.log(10) / mu ** 2) * np.arange(-mu, mu) ** 2)
waf = np.zeros((2 * mu, n_voices))
tfr = np.zeros((n_voices, tcol))
S1 = S2 = np.zeros((2 * n_voices,), dtype=complex)
mx1 = mx2 = np.zeros((2 * n_voices, 2 * mu))
for ti in range(tcol):
S1[:n_voices] = matxte1[:, ti]
mellin1 = np.fft.fftshift(np.fft.ifft(S1))
mx1 = l1 * mellin1.reshape(1, mellin1.shape[0]).repeat(2 * mu, 0)
mx1 = np.fft.fft(mx1, axis=0)
tx1 = mx1[:n_voices, :].T
S2[:n_voices] = matxte2[:, ti]
mellin2 = np.fft.fftshift(np.fft.ifft(S2))
mx2 = l2 * mellin2.reshape(1, mellin2.shape[0]).repeat(2 * mu, 0)
mx2 = np.fft.fft(mx2, axis=0)
tx2 = mx2[:n_voices, :].T
waf = np.real(tx1 * np.conj(tx2)) * G.reshape(G.shape[0], 1).repeat(n_voices, axis=1)
tfr[:, ti] = np.sum(waf) * geo_f
t = timestamps
f = geo_f
# Normalization
sp1 = np.fft.fft(hilbert(s1))
sp2 = np.fft.fft(hilbert(s2))
indmin = 1 + np.round(fmin * (xrow - 2))
indmax = 1 + np.round(fmax * (xrow - 2))
sp1_ana = sp1[indmin:(indmax + 1)]
sp2_ana = sp2[indmin:(indmax + 1)]
xx = np.dot(np.real(sp1_ana), np.real(sp2_ana))
xx += np.dot(np.imag(sp1_ana), np.imag(sp2_ana))
tfr = tfr * xx / integrate_2d(tfr, t, f) / n_voices
return tfr, t, f
def wide_band(signal, fmin=None, fmax=None, N=None):
if 1 in signal.shape:
signal = signal.ravel()
elif signal.ndim != 1:
raise ValueError("The input signal should be one dimensional.")
s_ana = hilbert(np.real(signal))
nx = signal.shape[0]
m = int(np.round(nx / 2.0))
t = np.arange(nx) - m
tmin = 0
tmax = nx - 1
T = tmax - tmin
# determine default values for fmin, fmax
if (fmin is None) or (fmax is None):
from matplotlib.mlab import find
STF = np.fft.fftshift(s_ana)
sp = np.abs(STF[:m])**2
maxsp = np.amax(sp)
f = np.linspace(0, 0.5, m + 1)
f = f[:m]
indmin = find(sp > maxsp / 100.0).min()
indmax = find(sp > maxsp / 100.0).max()
if fmin is None:
fmin = max([0.01, 0.05 * np.fix(f[indmin] / 0.05)])
if fmax is None:
fmax = 0.05 * np.ceil(f[indmax] / 0.05)
B = fmax - fmin
R = B / ((fmin + fmax) / 2.0)
nq = np.ceil((B * T * (1 + 2.0 / R) * np.log(
(1 + R / 2.0) / (1 - R / 2.0))) / 2.0)
nmin = nq - (nq % 2)
if N is None:
N = int(2**(nextpow2(nmin)))
# geometric sampling for the analyzed spectrum
k = np.arange(1, N + 1)
q = (fmax / fmin)**(1.0 / (N - 1))
geo_f = fmin * (np.exp((k - 1) * np.log(q)))
tfmatx = -2j * np.dot(t.reshape(-1, 1), geo_f.reshape(1, -1)) * np.pi
tfmatx = np.exp(tfmatx)
S = np.dot(s_ana.reshape(1, -1), tfmatx)
S = np.tile(S, (nx, 1))
Sb = S * tfmatx
tau = t
S = np.c_[S, np.zeros((nx, N))].T
Sb = np.c_[Sb, np.zeros((nx, N))].T
# mellin transform computation of the analyzed signal
p = np.arange(2 * N)
coef = np.exp(p / 2.0 * np.log(q))
mellinS = np.fft.fftshift(np.fft.ifft(S[:, 0] * coef))
mellinS = np.tile(mellinS, (nx, 1)).T
mellinSb = np.zeros((2 * N, nx), dtype=complex)
for i in range(nx):
mellinSb[:, i] = np.fft.fftshift(np.fft.ifft(Sb[:, i] * coef))
k = np.arange(1, 2 * N + 1)
scale = np.logspace(np.log10(fmin / fmax), np.log10(fmax / fmin), N)
theta = np.log(scale)
mellinSSb = mellinS * np.conj(mellinSb)
waf = np.fft.ifft(mellinSSb, N, axis=0)
no2 = int((N + N % 2) / 2.0)
waf = np.r_[waf[no2:(N + 1), :], waf[:no2, :]]
# normalization
s = np.real(s_ana)
SP = np.fft.fft(hilbert(s))
indmin = int(1 + np.round(fmin * (nx - 2)))
indmax = int(1 + np.round(fmax * (nx - 2)))
sp_ana = SP[(indmin - 1):indmax]
waf *= (np.linalg.norm(sp_ana)**2) / waf[no2 - 1, m - 1] / N
return waf, tau, theta
def _get_nvoices(self):
nq = np.ceil((self.bw * self.T * (1 + 2.0 / self.R) * np.log((1 + self.R / 2.0) / (1 - self.R / 2.0))) / 2)
nmin = nq - nq % 2
ndflt = 2 ** nextpow2(nmin)
self.n_voices = int(ndflt)