Skip to content

bob.ap
bob.ap
Commit 7c19c027 authored by André Anjos's avatar André Anjos 💬
Browse files
Options

Clean-up testing code a little bit

parent 8b5d813a
Hide whitespace changes
Inline Side-by-side
Showing with 144 additions and 363 deletions
xbob/ap/test_ceps.py
View file @ 7c19c027
......@@ -4,118 +4,17 @@
#
# Copyright (C) 2011-2013 Idiap Research Institute, Martigny, Switzerland
import os, sys
import os
import numpy
import array
import math
import time
import nose.tools
import pkg_resources
from . import Ceps
#############################################################################
# Tests blitz-based extrapolation implementation with values returned
#############################################################################
########################## Values used for the computation ##################
eps = 1e-3
#############################################################################
numpy.set_printoptions(precision=2, threshold=numpy.nan, linewidth=200)
def _read(filename):
"""Read video.FrameContainer containing preprocessed frames"""
fileName, fileExtension = os.path.splitext(filename)
wav_filename = filename
import scipy.io.wavfile
rate, data = scipy.io.wavfile.read(str(wav_filename)) # the data is read in its native format
if data.dtype =='int16':
data = numpy.cast['float'](data)
return [rate,data]
def compare(v1, v2, width):
return abs(v1-v2) <= width
def mel_python(f):
import math
return 2595.0*math.log10(1.+f/700.0)
def mel_inv_python(value):
return 700.0 * (10 ** (value / 2595.0) - 1)
from xbob.sp import fft
def sig_norm(win_length, frame, flag):
gain = 0.0
for i in range(win_length):
gain = gain + frame[i] * frame[i]
ENERGY_FLOOR = 1.0
if gain < ENERGY_FLOOR:
gain = math.log(ENERGY_FLOOR)
else:
gain = math.log(gain)
if(flag and gain != 0.0):
for i in range(win_length):
frame[i] = frame[i] / gain
return gain
def pre_emphasis(frame, win_length, a):
if (a < 0.0) or (a >= 1.0):
print("Error: The emphasis coeff. should be between 0 and 1")
if (a == 0.0):
return frame
else:
for i in range(win_length - 1, 0, -1):
frame[i] = frame[i] - a * frame[i - 1]
frame[0] = (1. - a) * frame[0]
return frame
def hamming_window(vector, hamming_kernel, win_length):
for i in range(win_length):
vector[i] = vector[i] * hamming_kernel[i]
return vector
def log_filter_bank(x, n_filters, p_index, win_size):
from xbob.sp import fft
x1 = numpy.array(x, dtype=numpy.complex128)
complex_ = fft(x1)
for i in range(0, int(win_size / 2) + 1):
re = complex_[i].real
im = complex_[i].imag
x[i] = math.sqrt(re * re + im * im)
filters = log_triangular_bank(x, n_filters, p_index)
return filters
def log_triangular_bank(data, n_filters, p_index):
a = 1.0 / (p_index[1:n_filters+2] - p_index[0:n_filters+1] + 1)
vec1 = list(numpy.arange(p_index[i], p_index[i + 1]) for i in range(0, n_filters))
vec2 = list(numpy.arange(p_index[i+1], p_index[i + 2] + 1) for i in range(0, n_filters))
res_ = numpy.array([(numpy.sum(data[vec1[i]]*(1.0 - a [i]* (p_index[i + 1]-(vec1[i])))) +
numpy.sum(data[vec2[i]] * (1.0 - a[i+1] * ( (vec2[i]) - p_index[i + 1]))))
for i in range(0, n_filters)])
FBANK_OUT_FLOOR = 1.0
filters = numpy.log(numpy.where(res_ < FBANK_OUT_FLOOR, FBANK_OUT_FLOOR, res_))
return filters
def dct_transform(filters, n_filters, dct_kernel, n_ceps, dct_norm):
if dct_norm:
dct_coeff = numpy.sqrt(2.0/(n_filters))
else :
dct_coeff = 1.0
ceps = numpy.zeros(n_ceps + 1)
vec = numpy.array(range(1, n_filters + 1))
for i in range(1, n_ceps + 1):
ceps[i - 1] = numpy.sum(filters[vec - 1] * dct_kernel[i - 1][0:n_filters])
ceps[i - 1] = ceps[i - 1] * dct_coeff
from . import Ceps
from .test_utils import *
return ceps
def cepstral_features_extraction(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win, pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
def cepstral_features_extraction(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
#########################
## Initialisation part ##
#########################
......@@ -230,7 +129,7 @@ def cepstral_features_extraction(rate_wavsample, win_length_ms, win_shift_ms, n_
# obj.assertAlmostEqual(frame[kk], f2[kk], 7, "Error in Pre-Emphasis Computation...")
f2=numpy.copy(frame)
filters = log_filter_bank(frame, n_filters, p_index, win_size)
filters, x = log_filter_bank(frame, n_filters, p_index, win_size)
#filt2 = ct.log_filter_bank(f2, win_size, n_filters)
......@@ -357,8 +256,7 @@ def cepstral_comparison_run(rate_wavsample, win_length_ms, win_shift_ms, n_filte
##################### Unit Tests ##################
def test_cepstral():
import pkg_resources
rate_wavsample = _read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
rate_wavsample = read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
win_length_ms = 20
win_shift_ms = 10
......
xbob/ap/test_energy.py
View file @ 7c19c027
......@@ -4,119 +4,16 @@
#
# Copyright (C) 2011-2013 Idiap Research Institute, Martigny, Switzerland
import os, sys
import unittest
import os
import numpy
import array
import math
import time
import pkg_resources
from . import Energy
from .test_utils import *
#############################################################################
# Tests blitz-based extrapolation implementation with values returned
#############################################################################
def energy_computation(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win, pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
########################## Values used for the computation ##################
eps = 1e-3
#############################################################################
numpy.set_printoptions(precision=2, threshold=numpy.nan, linewidth=200)
def _read(filename):
"""Read video.FrameContainer containing preprocessed frames"""
fileName, fileExtension = os.path.splitext(filename)
wav_filename = filename
import scipy.io.wavfile
rate, data = scipy.io.wavfile.read(str(wav_filename)) # the data is read in its native format
if data.dtype =='int16':
data = numpy.cast['float'](data)
return [rate,data]
def compare(v1, v2, width):
return abs(v1-v2) <= width
def mel_python(f):
import math
return 2595.0*math.log10(1.+f/700.0)
def mel_inv_python(value):
return 700.0 * (10 ** (value / 2595.0) - 1)
def sig_norm(win_length, frame, flag):
gain = 0.0
for i in range(win_length):
gain = gain + frame[i] * frame[i]
ENERGY_FLOOR = 1.0
if gain < ENERGY_FLOOR:
gain = math.log(ENERGY_FLOOR)
else:
gain = math.log(gain)
if(flag and gain != 0.0):
for i in range(win_length):
frame[i] = frame[i] / gain
return gain
def pre_emphasis(frame, win_length, a):
if (a < 0.0) or (a >= 1.0):
print("Error: The emphasis coeff. should be between 0 and 1")
if (a == 0.0):
return frame
else:
for i in range(win_length - 1, 0, -1):
frame[i] = frame[i] - a * frame[i - 1]
frame[0] = (1. - a) * frame[0]
return frame
def hamming_window(vector, hamming_kernel, win_length):
for i in range(win_length):
vector[i] = vector[i] * hamming_kernel[i]
return vector
def log_filter_bank(x, n_filters, p_index, win_size):
from xbob.sp import fft
x1 = numpy.array(x, dtype=numpy.complex128)
complex_ = fft(x1)
for i in range(0, win_size / 2 + 1):
re = complex_[i].real
im = complex_[i].imag
x[i] = math.sqrt(re * re + im * im)
filters = log_triangular_bank(x, n_filters, p_index)
return filters, x
def log_triangular_bank(data, n_filters, p_index):
a = 1.0 / (p_index[1:n_filters+2] - p_index[0:n_filters+1] + 1)
vec1 = list(numpy.arange(p_index[i], p_index[i + 1]) for i in range(0, n_filters))
vec2 = list(numpy.arange(p_index[i+1], p_index[i + 2] + 1) for i in range(0, n_filters))
res_ = numpy.array([(numpy.sum(data[vec1[i]]*(1.0 - a [i]* (p_index[i + 1]-(vec1[i])))) +
numpy.sum(data[vec2[i]] * (1.0 - a[i+1] * ( (vec2[i]) - p_index[i + 1]))))
for i in range(0, n_filters)])
FBANK_OUT_FLOOR = 1.0
filters = numpy.log(numpy.where(res_ < FBANK_OUT_FLOOR, FBANK_OUT_FLOOR, res_))
return filters
def dct_transform(filters, n_filters, dct_kernel, n_ceps, dct_norm):
if dct_norm:
dct_coeff = numpy.sqrt(2.0/(n_filters))
else :
dct_coeff = 1.0
ceps = numpy.zeros(n_ceps + 1)
vec = numpy.array(range(1, n_filters + 1))
for i in range(1, n_ceps + 1):
ceps[i - 1] = numpy.sum(filters[vec - 1] * dct_kernel[i - 1][0:n_filters])
ceps[i - 1] = ceps[i - 1] * dct_coeff
return ceps
def energy_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
#########################
## Initialisation part ##
#########################
......@@ -203,40 +100,35 @@ def energy_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filte
return data
def energy_comparison_run(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
def energy_comparison_run(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win, pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta):
c = Energy(rate_wavsample[0], win_length_ms, win_shift_ms)
#ct = TestCeps(c)
A = c(rate_wavsample[1])
B = energy_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm,
B = energy_computation(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm,
f_min, f_max, delta_win, pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta)
diff=numpy.sum(numpy.sum((A-B)*(A-B)))
obj.assertAlmostEqual(diff, 0., 7, "Error in Ceps Analysis")
assert numpy.allclose(diff, 0., rtol=1e-07, atol=1e-05)
##################### Unit Tests ##################
class EnergyTest(unittest.TestCase):
"""Test the Energy extraction"""
def test_energy(self):
import pkg_resources
rate_wavsample = _read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
win_length_ms = 20
win_shift_ms = 10
n_filters = 24
n_ceps = 19
f_min = 0.
f_max = 4000.
delta_win = 2
pre_emphasis_coef = 0.97
dct_norm = True
mel_scale = True
with_energy = True
with_delta = True
with_delta_delta = True
energy_comparison_run(self,rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta)
def test_energy():
rate_wavsample = read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
win_length_ms = 20
win_shift_ms = 10
n_filters = 24
n_ceps = 19
f_min = 0.
f_max = 4000.
delta_win = 2
pre_emphasis_coef = 0.97
dct_norm = True
mel_scale = True
with_energy = True
with_delta = True
with_delta_delta = True
energy_comparison_run(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win, pre_emphasis_coef, mel_scale, with_energy, with_delta, with_delta_delta)
xbob/ap/test_spectrogram.py
View file @ 7c19c027
......@@ -4,118 +4,15 @@
#
# Copyright (C) 2011-2013 Idiap Research Institute, Martigny, Switzerland
import os, sys
import unittest
import os
import numpy
import array
import math
import time
import pkg_resources
from . import Spectrogram
from .test_utils import *
#############################################################################
# Tests blitz-based extrapolation implementation with values returned
#############################################################################
########################## Values used for the computation ##################
eps = 1e-3
#############################################################################
numpy.set_printoptions(precision=2, threshold=numpy.nan, linewidth=200)
def _read(filename):
"""Read video.FrameContainer containing preprocessed frames"""
fileName, fileExtension = os.path.splitext(filename)
wav_filename = filename
import scipy.io.wavfile
rate, data = scipy.io.wavfile.read(str(wav_filename)) # the data is read in its native format
if data.dtype =='int16':
data = numpy.cast['float'](data)
return [rate,data]
def compare(v1, v2, width):
return abs(v1-v2) <= width
def mel_python(f):
import math
return 2595.0*math.log10(1.+f/700.0)
def mel_inv_python(value):
return 700.0 * (10 ** (value / 2595.0) - 1)
def sig_norm(win_length, frame, flag):
gain = 0.0
for i in range(win_length):
gain = gain + frame[i] * frame[i]
ENERGY_FLOOR = 1.0
if gain < ENERGY_FLOOR:
gain = math.log(ENERGY_FLOOR)
else:
gain = math.log(gain)
if(flag and gain != 0.0):
for i in range(win_length):
frame[i] = frame[i] / gain
return gain
def pre_emphasis(frame, win_length, a):
if (a < 0.0) or (a >= 1.0):
print("Error: The emphasis coeff. should be between 0 and 1")
if (a == 0.0):
return frame
else:
for i in range(win_length - 1, 0, -1):
frame[i] = frame[i] - a * frame[i - 1]
frame[0] = (1. - a) * frame[0]
return frame
def hamming_window(vector, hamming_kernel, win_length):
for i in range(win_length):
vector[i] = vector[i] * hamming_kernel[i]
return vector
def log_filter_bank(x, n_filters, p_index, win_size):
from xbob.sp import fft
x1 = numpy.array(x, dtype=numpy.complex128)
complex_ = fft(x1)
for i in range(0, int(win_size / 2) + 1):
re = complex_[i].real
im = complex_[i].imag
x[i] = math.sqrt(re * re + im * im)
filters = log_triangular_bank(x, n_filters, p_index)
return filters, x
def log_triangular_bank(data, n_filters, p_index):
a = 1.0 / (p_index[1:n_filters+2] - p_index[0:n_filters+1] + 1)
vec1 = list(numpy.arange(p_index[i], p_index[i + 1]) for i in range(0, n_filters))
vec2 = list(numpy.arange(p_index[i+1], p_index[i + 2] + 1) for i in range(0, n_filters))
res_ = numpy.array([(numpy.sum(data[vec1[i]]*(1.0 - a [i]* (p_index[i + 1]-(vec1[i])))) +
numpy.sum(data[vec2[i]] * (1.0 - a[i+1] * ( (vec2[i]) - p_index[i + 1]))))
for i in range(0, n_filters)])
FBANK_OUT_FLOOR = 1.0
filters = numpy.log(numpy.where(res_ < FBANK_OUT_FLOOR, FBANK_OUT_FLOOR, res_))
return filters
def dct_transform(filters, n_filters, dct_kernel, n_ceps, dct_norm):
if dct_norm:
dct_coeff = numpy.sqrt(2.0/(n_filters))
else :
dct_coeff = 1.0
ceps = numpy.zeros(n_ceps + 1)
vec = numpy.array(range(1, n_filters + 1))
for i in range(1, n_ceps + 1):
ceps[i - 1] = numpy.sum(filters[vec - 1] * dct_kernel[i - 1][0:n_filters])
ceps[i - 1] = ceps[i - 1] * dct_coeff
return ceps
def spectrogram_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, f_min, f_max,
pre_emphasis_coef, mel_scale):
def spectrogram_computation(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, f_min, f_max, pre_emphasis_coef, mel_scale):
#########################
## Initialisation part ##
#########################
......@@ -215,24 +112,20 @@ def spectrogram_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_
return data
def spectrogram_comparison_run(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
def spectrogram_comparison_run(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale):
c = Spectrogram(rate_wavsample[0], win_length_ms, win_shift_ms, n_filters, f_min, f_max, pre_emphasis_coef, mel_scale)
A = c(rate_wavsample[1])
B = spectrogram_computation(obj, rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps,
f_min, f_max, pre_emphasis_coef, mel_scale)
B = spectrogram_computation(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, f_min, f_max, pre_emphasis_coef, mel_scale)
diff=numpy.sum(numpy.sum((A-B)*(A-B)))
obj.assertAlmostEqual(diff, 0., 7, "Error in Ceps Analysis")
assert numpy.allclose(diff, 0., rtol=1e-07, atol=1e-05)
##################### Unit Tests ##################
class SpecTest(unittest.TestCase):
"""Test the Spectral feature extraction"""
def test_spectrogram():
def test_spectrogram(self):
import pkg_resources
rate_wavsample = _read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
rate_wavsample = read(pkg_resources.resource_filename(__name__, os.path.join('data', 'sample.wav')))
win_length_ms = 20
win_shift_ms = 10
......@@ -244,6 +137,4 @@ class SpecTest(unittest.TestCase):
pre_emphasis_coef = 0.97
dct_norm = True
mel_scale = True
spectrogram_comparison_run(self,rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win,
pre_emphasis_coef, mel_scale)
spectrogram_comparison_run(rate_wavsample, win_length_ms, win_shift_ms, n_filters, n_ceps, dct_norm, f_min, f_max, delta_win, pre_emphasis_coef, mel_scale)
xbob/ap/test_utils.py 0 → 100644
View file @ 7c19c027
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :
# Elie Khoury <Elie.Khoury@idiap.ch>
# Andre Anjos <andre.anjos@idiap.ch>
#
# Copyright (C) 2011-2014 Idiap Research Institute, Martigny, Switzerland
import os
import math
import numpy
import scipy.io.wavfile
from xbob.sp import fft
def read(filename):
"""Read video.FrameContainer containing preprocessed frames"""
fileName, fileExtension = os.path.splitext(filename)
wav_filename = filename
rate, data = scipy.io.wavfile.read(str(wav_filename)) # the data is read in its native format
if data.dtype =='int16':
data = numpy.cast['float'](data)
return [rate,data]
def compare(v1, v2, width):
return abs(v1-v2) <= width
def mel_python(f):
return 2595.0*math.log10(1.+f/700.0)
def mel_inv_python(value):
return 700.0 * (10 ** (value / 2595.0) - 1)
def sig_norm(win_length, frame, flag):
gain = 0.0
for i in range(win_length):
gain = gain + frame[i] * frame[i]
ENERGY_FLOOR = 1.0
if gain < ENERGY_FLOOR:
gain = math.log(ENERGY_FLOOR)
else:
gain = math.log(gain)
if(flag and gain != 0.0):
for i in range(win_length):
frame[i] = frame[i] / gain
return gain
def pre_emphasis(frame, win_length, a):
if (a < 0.0) or (a >= 1.0):
print("Error: The emphasis coeff. should be between 0 and 1")
if (a == 0.0):
return frame
else:
for i in range(win_length - 1, 0, -1):
frame[i] = frame[i] - a * frame[i - 1]
frame[0] = (1. - a) * frame[0]
return frame
def hamming_window(vector, hamming_kernel, win_length):
for i in range(win_length):
vector[i] = vector[i] * hamming_kernel[i]
return vector
def log_filter_bank(x, n_filters, p_index, win_size):
x1 = numpy.array(x, dtype=numpy.complex128)
complex_ = fft(x1)
for i in range(0, int(win_size / 2) + 1):
re = complex_[i].real
im = complex_[i].imag
x[i] = math.sqrt(re * re + im * im)
filters = log_triangular_bank(x, n_filters, p_index)
return filters, x
def log_triangular_bank(data, n_filters, p_index):
a = 1.0 / (p_index[1:n_filters+2] - p_index[0:n_filters+1] + 1)
vec1 = list(numpy.arange(p_index[i], p_index[i + 1]) for i in range(0, n_filters))
vec2 = list(numpy.arange(p_index[i+1], p_index[i + 2] + 1) for i in range(0, n_filters))
res_ = numpy.array([(numpy.sum(data[vec1[i]]*(1.0 - a [i]* (p_index[i + 1]-(vec1[i])))) +
numpy.sum(data[vec2[i]] * (1.0 - a[i+1] * ( (vec2[i]) - p_index[i + 1]))))
for i in range(0, n_filters)])
FBANK_OUT_FLOOR = 1.0
filters = numpy.log(numpy.where(res_ < FBANK_OUT_FLOOR, FBANK_OUT_FLOOR, res_))
return filters
def dct_transform(filters, n_filters, dct_kernel, n_ceps, dct_norm):
if dct_norm:
dct_coeff = numpy.sqrt(2.0/(n_filters))
else :
dct_coeff = 1.0
ceps = numpy.zeros(n_ceps + 1)
vec = numpy.array(range(1, n_filters + 1))
for i in range(1, n_ceps + 1):
ceps[i - 1] = numpy.sum(filters[vec - 1] * dct_kernel[i - 1][0:n_filters])
ceps[i - 1] = ceps[i - 1] * dct_coeff
return ceps
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment