def __init__(
self,
parent,
dump_vectors_beyond=20,
lambda_border=0,
two_dim_mode='separate',
reinit=False,
):
Processor.__init__(self,
parent,
Default_Preprocessor=Cepstrogram,
reinit=reinit)
input_vectors = self.parent_processor.feature_vectors
input_shape = shape(input_vectors)
if len(input_shape) > 2:
if False and two_dim_mode == 'separate':
print "mode:", two_dim_mode
keep_from_band_variations = 3
svd_input_vectors = zeros(
(input_shape[0],
input_shape[1] * keep_from_band_variations), Float)
counter = 0
for spectrum_of_band in parent.feature_vectors:
print counter
svd_input_vectors[:, counter : counter + keep_from_band_variations - 1] = \
svdeofs.svdeofs(input_vectors)[0][:, :keep_from_band_variations]
counter += keep_from_band_variations
input_vectors = svd_input_vectors
else:
input_vectors = reshape(
input_vectors,
(input_shape[0], input_shape[1] * input_shape[2]))
print shape(input_vectors)
(z, self.lambdas, self.EOFs) = svdeofs.svdeofs(input_vectors)
if not dump_vectors_beyond:
if lambda_border:
dump_vectors_beyond = argmin(
greater(self.lambdas, lambda_border))
if dump_vectors_beyond:
self.feature_vectors = z[:, :dump_vectors_beyond]
else:
self.feature_vectors = z
def __init__(self, parent, framesize = None, hopsize = 100, number_of_vectors_used = 15): if not framesize: lp = len(parent) #logfs = min(int(scipy.log2(lp / 32)), 10) log2 = lambda x: math.log(x) / math.log(2) logfs = min(int(log2(lp / 32)), 10) framesize = pow(2, logfs) print framesize if hopsize > framesize / 4: hopsize = framesize / 4 self.hopsize = hopsize self.framesize = framesize #if len(parent) < framesize: # self._delete_from_parents() # raise UnderflowError, "Thingy is too small...\nHere I pull the plug in order to avoid a segfaulty thingy." input_array = parent self.feature_vectors = None interesting_parts = input_array[:, :number_of_vectors_used] interesting_parts = numarray.transpose(interesting_parts) for row in interesting_parts: specspectrum = calculate_spectrogram(row, framesize = framesize, hopsize = hopsize) import cPickle z, lambdas, EOFs = svdeofs.svdeofs(specspectrum) print ".", s = z[:, :15] svd_fft_fft_vectors = numarray.transpose(s) #print "svd_eofs ok!" if self.feature_vectors is None: self.feature_vectors = numarray.transpose(svd_fft_fft_vectors) else: self.feature_vectors = numarray.concatenate((self.feature_vectors, numarray.transpose(svd_fft_fft_vectors)), 1) #print "svd_fft_fft_vectors shapes:", shape(svd_fft_fft_vectors), shape(self.feature_vectors) z, lambdas, EOFs = svdeofs.svdeofs(self.feature_vectors) self.feature_vectors = z[:, :15] self.arr = self.feature_vectors
def svd_eofs(feature_vectors, number_of_vectors_to_keep=15, show_lambdas=0):
# Should be using svd_eofs.SVDEOFs here
z, lambdas, EOFs = svdeofs.svdeofs(feature_vectors)
if show_lambdas:
my_show.show(lambdas)
return (z[:, :number_of_vectors_to_keep], lambdas)
def __init__(
self,
parent,
framesize=None,
hopsize=100,
window_function=signal.hanning,
number_of_vectors_used=15,
reinit=False,
):
Processor.__init__(
self,
parent,
# Default_Preprocessor = Cepstrogram,
Default_Preprocessor=Spectrogram,
reinit=reinit)
if not framesize:
lp = len(self.parent_processor)
logfs = min(int(log2(lp / 32)), 10)
framesize = pow(2, logfs)
if hopsize > framesize / 16:
hopsize = framesize / 16
print "hopsize: ", hopsize
self._set_samplerate(hopsize)
self.hopsize = hopsize
self.framesize = framesize
print "EOF..."
if len(self.parent_processor) < framesize:
self._delete_from_parents()
raise UnderflowError, "Thingy is too small...\nHere I pull the plug in order to avoid a segfaulty thingy."
self.svd_fft_fft(self.parent_processor.feature_vectors,
framesize=framesize,
hopsize=hopsize,
window_function=window_function,
number_of_vectors_used=number_of_vectors_used)
z, lambdas, EOFs = svdeofs.svdeofs(self.feature_vectors)
self.feature_vectors = z[:, :15]
#self.feature_vectors, self.lambdas = svd_eofs(self.feature_vectors,
# number_of_vectors_to_keep = 15)
Segmentations.Frames(
self,
len(self.feature_vectors),
self.parent_processor.samplerate,
framesize,
hopsize,
window_function=window_function,
)
def __init__(
self,
parent,
framesize=32,
hopsize=8,
window_function=signal.hanning,
number_of_vectors_used=15,
reinit=False,
):
window_function = signal.hanning
Processor.__init__(
self,
parent,
# Default_Preprocessor = Cepstrogram,
Default_Preprocessor=Spectrogram,
reinit=reinit)
self._set_samplerate(hopsize)
self.hopsize = hopsize
self.framesize = framesize
print "EOF..."
if len(self.parent_processor) < framesize:
self._delete_from_parents()
raise UnderflowError, "Thingy is too small...\nHere I pull the plug in order to avoid a segfaulty thingy."
self.svd_fft_fft(self.parent_processor.feature_vectors,
framesize=framesize,
hopsize=hopsize,
window_function=window_function,
number_of_vectors_used=number_of_vectors_used)
self.test = self.feature_vectors
z, lambdas, EOFs = svdeofs.svdeofs(self.feature_vectors)
self.feature_vectors = z[:, :15]
Segmentations.Frames(
self,
len(self.feature_vectors),
self.parent_processor.samplerate,
framesize,
hopsize,
window_function=window_function,
)
def ceof_scalar2D(data):
""" Estimate the complex EOF on a 2D array.
Time should be the first dimension, so that the PC (eigenvalues) will
be in respect to the first dimension.
"""
assert type(data) is np.ndarray, \
"ceof_scalar2D requires an ndarray but got: %s" % type(data)
assert np.isfinite(data).all(), \
"ceof_scalar2D requires a full valid values array"
# ---- Creating the complex field using Hilbert transform
input_H = numpy.empty(data.shape, dtype=data.dtype)
for i in range(data.shape[1]):
input_H[:, i] = scipy.fftpack.hilbert(data[:, i])
U = data + 1j * input_H
pcs, lambdas, eofs = svdeofs(U)
return pcs, lambdas, eofs
def ceof_scalar2D(data):
""" Estimate the complex EOF on a 2D array.
Time should be the first dimension, so that the PC (eigenvalues) will
be in respect to the first dimension.
"""
assert type(data) is np.ndarray, \
"ceof_scalar2D requires an ndarray but got: %s" % type(data)
assert np.isfinite(data).all(), \
"ceof_scalar2D requires a full valid values array"
# ---- Creating the complex field using Hilbert transform
input_H = numpy.empty(data.shape, dtype=data.dtype)
for i in range(data.shape[1]):
input_H[:,i] = scipy.fftpack.hilbert(data[:,i])
U = data + 1j*input_H
pcs, lambdas, eofs = svdeofs(U)
return pcs, lambdas, eofs