def calculate_spectrogram(input_array, framesize, hopsize, window_function, keep_bands_until = 0, axis = 0, show_progressbar = 0 ): # TODO Segmentation should be determined here! #Segmentations.Frames(self, # len(self. # print "Calculating FFT..." do_fft = lambda arr: abs(FFT.real_fft(arr, axis = axis))[:keep_bands_until] keep_bands_until = keep_bands_until or int(framesize / 2) input_shape = map(None, shape(input_array)) window = window_function(framesize) if len(input_shape) > 1: window = transpose(array([list(window)] * input_shape[1])) # print "input_shape:", shape(input_array) zeros_shape = input_shape # this allows for both 1 and 2 dim inputs zeros_shape[0] = framesize / 2 input_array_plus_zeros = concatenate((zeros(zeros_shape), input_array, zeros(zeros_shape))) fft_range = range(0, len(input_array) - framesize, hopsize) fft_array = zeros(([len(fft_range)] + map(None, shape(do_fft(window)))), # do_fft(window) is used here because it gives the right shape Float32) * 1.0 # this *1.0 is necessary! # print shape(fft_array) if show_progressbar: pbar = Progressbar(len(fft_range)) output_counter = 0 for i in fft_range: frame = window * input_array_plus_zeros[i : i + framesize] fft_array[output_counter] = 10 * log10(0.1 + do_fft(frame)) output_counter += 1 if show_progressbar: divmod(output_counter, 128)[1] or pbar.update(output_counter) if show_progressbar: pbar.finish() # print "outputshape:", shape(fft_array) return fft_array
def OLD_calc_fft(input_array,
hopsize,
framesize,
window_function=signal.hanning):
keep_bands_until = keep_bands_until or int(framesize / 2)
do_fft = lambda arr: abs(FFT.real_fft(arr))[:keep_bands_until]
input_shape = map(None, shape(input_array))
window = window_function(framesize)
if len(input_shape) > 1:
window = transpose(array([list(window)] * input_shape[1]))
print "input_shape:", shape(input_array)
zeros_shape = input_shape # this allows for both 1 and 2 dim inputs
zeros_shape[0] = framesize
input_array_plus_zeros = concatenate((input_array, zeros(zeros_shape)))
fft_range = range(0, len(input_array) - framesize, hopsize)
if show_progressbar:
pbar = Progressbar(len(fft_range))
output_counter = 0
fft_array = zeros(([len(fft_range)] + map(None, shape(do_fft(window)))),
Float32) * 1.0 # this *1.0 is necessary!
print shape(fft_array)
for i in fft_range:
frame = window * input_array_plus_zeros[i:i + framesize]
fft_array[output_counter] = 10 * log10(0.1 + do_fft(frame))
if show_progressbar:
divmod(output_counter, 128)[1] or pbar.update(output_counter)
output_counter += 1
if show_progressbar:
pbar.finish()
return fft_array
def svd_fft_fft(self,
input_array,
framesize,
hopsize,
window_function=signal.hanning,
number_of_vectors_used=256):
pbar = Progressbar(number_of_vectors_used)
self.feature_vectors = None
interesting_parts = input_array[:, :number_of_vectors_used]
interesting_parts = transpose(interesting_parts)
for row in interesting_parts:
# print "row shape:", shape(row)
pbar.update()
#fft = parent.feature_vectors
specspectrum = calculate_spectrogram(
row,
framesize=framesize,
hopsize=hopsize,
window_function=window_function,
show_progressbar=0)
#z, lambdas, EOFs = svdeofs.svdeofs(fft)
#svd_fft_fft_vectors = transpose(z[:, :15])
svd_fft_fft_vectors = transpose(svd_eofs(specspectrum)[0])
if self.feature_vectors is None:
self.feature_vectors = transpose(svd_fft_fft_vectors)
else:
self.feature_vectors = concatenate(
(self.feature_vectors, transpose(svd_fft_fft_vectors)), 1)
#print "svd_fft_fft_vectors shapes:", shape(svd_fft_fft_vectors), shape(self.feature_vectors)
pbar.finish()
def different_kernel_sizes(self, arr, l = 20): a = zeros((l, len(arr)), Float) pb = Progressbar(l) print "b" for i in range(0, l): # a[l - 1 - i] = diag(sim_matrix_find_segments(arr, i, 2)) a[l - 1 - i] = diag(sim_matrix_find_segments(arr, i, 2)) print "b" pb.update() print "c" pb.finish() return a
def OLD_vertical_structure_filter(arr, delta = 10, boundary = 10): new_arr = copy.copy(arr) arr_binary = greater(arr, 0) l = len(arr) pb = Progressbar(l) for x in range(l): pb.update() for y in range(delta, l - x - delta): if sum(arr_binary[y - delta : y + delta, x]) < boundary: new_arr[y, x] = 0 pb.finish() return new_arr
def OLD_calculate_distance_matrix(self): """Calculate the similarity matrix.""" W = Numeric.array(self.feature_vectors) length = W.shape[0] S = zeros((length, length)) * 1.0 pbar = Progressbar(length) for i in range(length): pbar.update() for j in range(i): S[j, i] = S[i, j] = self.distance_function(W[i], W[j]) pbar.finish() return S
def window_for_window(self, window): arr = zeros((self.parent_segmentation.number_of_frames), Float) a = 0 b = len(window) w = self.window_function(self.framesize) w *= self.hopsize / sum(w) pb = Progressbar(len(window)) for i in range(len(window)): # print len(arr[i * self.hopsize : (i + 1) * self.hopsize - 1]), len(window[i] * w) pb.update() part = arr[i * self.hopsize : i * self.hopsize + self.framesize] part += (window[i] * w)[:len(part)] pb.finish() # assert alltrue(greater_equal(arr, 0)) return arr
def __init__(self, parent_processor, parent_borders): parent = parent_processor Parts.__init__(self, parent) self.parent_borders = parent_borders bp = parent_borders.get_border_points() l = [] pb = Progressbar(len(bp)) for cur_pos, parent_weight in bp: start, end = parent_borders.segmentation.parent_segments(cur_pos) sv = sum(calculate_distance_band(parent_processor.feature_vectors[start : end], 150)) # MAYBE TODO: sum -> cumsum # MAYBE TODO: multiply sv with window_function(framesize)... am = argmax(sv) own_weight = sqrt(sv[am] / 1024 / 50) b = MorePreciseBorder(self, start + am, parent_weight, own_weight) l.append([start + am, parent_weight, own_weight]) self.append(b) pb.update() pb.finish() self.feature_vectors = array(l)