def rx(samples, prototype, osr):
# 0. prep
coeffs = _polyphase_filter_coeffs(prototype[::-1], osr)
# print coeffs
# 1. input commutator
if coeffs.size / 2 == len(prototype) - 1:
# PHYDYAS filter: first and last sample is zero and therefore cut
# --> samples must be cut as well
samples = flipud(samples[1:].reshape((-1, osr // 2)).T)
else:
# other filters use the full OSR*overlap + 1 samples
# --> prefix some zeros to also get transient behavior right
samples = flipud(append(
zeros(osr - 1, dtype=samples[0].dtype), samples).reshape((-1, osr // 2)).T)
# print samples
# 2. polyphase filters
samples = append(samples, empty_like(samples), axis=0)
# out = empty((samples.shape[0], samples.shape[1] + coeffs.shape[1] - 1), dtype=samples.dtype)
out = samples[:, :-(coeffs.shape[1] - 1)]
# print np.round(samples)
# reverse iteration to not override samples for l >= osr/2
for l in reversed(range(osr)):
# run filter (draw samples from the first half for l >= osr/2
out[l, :] = convolve(samples[l % (osr // 2), :], coeffs[l, :], 'valid')
# print "\n", np.round(samples)
# print "out\n", np.round(out)
# 3. spin polyphase signals
out[:] = osr * ifft(out, osr, 0)
return out
def b_conv(v1, v2):
w1 = dual.fft(v1[rand_perm[0]])
w2 = dual.fft(v2[rand_perm[1]])
return np.real_if_close(dual.ifft(w1 * w2))
def project_H(self, H, copy=False, phase=False):
Hft = fft(H, axis=1)
if phase:
Hsum = np.sum(Hft, axis=1)
Hphase = Hsum / np.abs(Hsum)
Hft /= Hphase[:, None]
Hft.imag = 0 # Keep only the real part
Hft[Hft.real <= 0] = 0
if copy:
H = ifft(Hft, axis=1)
else:
H[:, :] = ifft(Hft, axis=1)
N = H.shape[1]
KKfactor = np.linspace(2, 0, N, endpoint=False)
KKfactor[0] = 1
H *= KKfactor[None, :]
if phase:
return H, Hphase
else:
return H
def __call__(self, realdata):
N = realdata.shape[0]
## Not sure which of these ways of calculating tweak_fid is best
## TODO: test them out more thoroughly
tweak_fid = np.empty((N, ), dtype=float)
tweak_fid[0] = 1
tweak_fid[1:] = np.linspace(2, 0, N - 1, endpoint=True)
#tweak_fid = np.linspace(2, 0, N, endpoint=False)
#tweak_fid[0] = 1
fid_guess = ifft(fftshift(realdata))
return fftshift(fft(fid_guess * tweak_fid))
def project_H(self, H, copy=False, phase=False):
Hft = fft(H, axis=1)
if phase:
Hsum = np.sum(Hft, axis=1)
Hphase = Hsum / np.abs(Hsum)
Hft /= Hphase[:, None]
Hft.imag = 0 # Keep only the real part
Hft[Hft.real <= 0] = 0
if copy:
H = ifft(Hft, axis=1)
else:
H[:, :] = ifft(Hft, axis=1)
N = H.shape[1]
KKfactor = np.linspace(2, 0, N, endpoint=False)
KKfactor[0] = 1
H *= KKfactor[None, :]
if phase:
return H, Hphase
else:
return H
def __call__(self, realdata):
N = realdata.shape[0]
## Not sure which of these ways of calculating tweak_fid is best
## TODO: test them out more thoroughly
tweak_fid = np.empty((N,), dtype=float)
tweak_fid[0] = 1
tweak_fid[1:] = np.linspace(2, 0, N - 1, endpoint=True)
#tweak_fid = np.linspace(2, 0, N, endpoint=False)
#tweak_fid[0] = 1
fid_guess = ifft(fftshift(realdata))
return fftshift(fft(fid_guess * tweak_fid))
def generateZadoffTimeDomainSeq(seqID):
fDat = sequences.zcsequence(seqID, 63)
# fDat is in the frequency domain, convert it to time domain
# This code was taken from JiaoXianjun/LTE-Cell-Scanner in tle_lib.cpp
idft_in = np.concatenate((np.zeros(1, np.complex64), fDat[31:61 + 1],
np.zeros(65, np.complex64), fDat[0:30 + 1]))
td = ifft(idft_in) * np.sqrt(128.0 / 62.0) * np.sqrt(128.)
return np.concatenate((td[119:127 + 1], td))
return td
def f(vector_list):
if len(vector_list) == 0:
return np.zeros(self.dimension)
elif len(vector_list) == 1:
return vector_list[0]
else:
v1 = dual.fft(left_permutation(f(vector_list[:-1])))
v2 = dual.fft(right_permutation(vector_list[len(vector_list)-1]))
return dual.ifft(v1 * v2)
def project_H(self, H, copy=False):
N = H.shape[1]
if copy:
H = np.copy(H)
H.imag = 0 # Keep only the real part
H.real[H.real <= 0] = 0
# Do the Kramers-Kronig convolution
Hft = ifft(H, axis=1)
KKfactor = np.linspace(2, 0, N, endpoint=False)
KKfactor[0] = 1
Hft *= KKfactor[None, :]
H = fft(Hft, axis=1)
# Zero out small negative numbers caused by rounding errors
H.real[H.real <= 0] = 0
return H
def project_H(self, H, copy=False):
N = H.shape[1]
if copy:
H = np.copy(H)
H.imag = 0 # Keep only the real part
H.real[H.real <= 0] = 0
# Do the Kramers-Kronig convolution
Hft = ifft(H, axis=1)
KKfactor = np.linspace(2, 0, N, endpoint=False)
KKfactor[0] = 1
Hft *= KKfactor[None, :]
H = fft(Hft, axis=1)
# Zero out small negative numbers caused by rounding errors
H.real[H.real <= 0] = 0
return H
def NMR_svd_initialise(X, n_components, constraint):
"""Calculate a starting point for the generalised NMF fit of NMR
Does a standard NNDSVD (as used in standard NMF) on the real part of the
(approximately phased) NMR spectrum. If the constraint is a FIDconstraint,
Fourier transforms the X array so the NNDSVD is done on the spectrum."""
if isinstance(constraint, FIDConstraint):
FT = True
else:
FT = False
if FT:
X = fft(X, axis=1)
W, H = realpart_svd_initialise(X, n_components)
if FT:
H = ifft(H, axis=1)
H = constraint.project_H(H) # Need to add back in the imaginary part
W, H = constraint.normalise(W, H)
return W, H
def NMR_svd_initialise(X, n_components, constraint):
"""Calculate a starting point for the generalised NMF fit of NMR
Does a standard NNDSVD (as used in standard NMF) on the real part of the
(approximately phased) NMR spectrum. If the constraint is a FIDconstraint,
Fourier transforms the X array so the NNDSVD is done on the spectrum."""
if isinstance(constraint, FIDConstraint):
FT = True
else:
FT = False
if FT:
X = fft(X, axis=1)
W, H = realpart_svd_initialise(X, n_components)
if FT:
H = ifft(H, axis=1)
H = constraint.project_H(H) # Need to add back in the imaginary part
W, H = constraint.normalise(W, H)
return W, H
def imag(self, realdata):
N = realdata.shape[0]
tweak_fid = np.linspace(-1j, 1j, N, endpoint=False)
tweak_fid[0] = 0
fid_guess = ifft(fftshift(realdata))
return fftshift(fft(fid_guess * tweak_fid)).real
def mifft(f):
if size(f,0)==1 or size(f,1)==1:
F = fftshift(ifft(ifftshift(f)))
else:
F = fftshift(ifft2(ifftshift(f)))
return F
def b_conv(v1, v2):
w1 = dual.fft(v1[rand_perm[0]])
w2 = dual.fft(v2[rand_perm[1]])
return np.real_if_close(dual.ifft(w1 * w2))
def imag(self, realdata):
N = realdata.shape[0]
tweak_fid = np.linspace(-1j, 1j, N, endpoint=False)
tweak_fid[0] = 0
fid_guess = ifft(fftshift(realdata))
return fftshift(fft(fid_guess * tweak_fid)).real
