def test_pgc2():
assert nb_channel==1
#~ in_buffer = hls.moving_sinus(length, sample_rate=sample_rate, speed = .5, f1=50., f2=2000., ampl = .8)
#~ in_buffer = hls.moving_erb_noise(length, sample_rate=sample_rate,)
in_buffer = hls.whitenoise(length, sample_rate=sample_rate,)
in_buffer = np.tile(in_buffer[:, None],(1, nb_channel))
loss_params = { 'left' : {'freqs' : [ 125*2**i for i in range(7) ], 'compression_degree': [0]*7, 'passive_loss_db' : [0]*7 } }
loss_params['right'] = loss_params['left']
processing_conf = dict(nb_freq_band=32, level_max=120, level_step=1, debug_mode=True,
low_freq = 60., high_freq = 15000.,
loss_params = loss_params,
chunksize=chunksize, backward_chunksize=backward_chunksize)
processing, online_arrs = hls.run_class_offline(hls.InvCGC, in_buffer, chunksize, sample_rate, processing_conf=processing_conf, buffersize_margin=backward_chunksize)
freq_band = 4
online_hpaf = online_arrs['hpaf']
online_pgc2 = online_arrs['pgc2']
offline_pgc2 = online_pgc2.copy()
n = processing.nb_freq_band
for i in range(n):
offline_pgc2[:, i] = scipy.signal.sosfilt(processing.coefficients_pgc[i, :,:], online_hpaf[::-1,i])[::-1]
online_pgc2 = online_pgc2[:-backward_chunksize]
offline_pgc2 = offline_pgc2[:-backward_chunksize]
residual = np.abs((online_pgc2.astype('float64')-offline_pgc2.astype('float64'))/np.mean(np.abs(offline_pgc2.astype('float64')), axis=0))
print(np.max(residual, axis=0))
print(np.max(residual))
#~ freq_band = 4
fig, ax = plt.subplots(nrows = 2, sharex=True)
#~ ax[0].plot(online_hpaf[:, freq_band], color = 'b')
ax[0].plot(offline_pgc2[:, freq_band], color = 'g')
ax[0].plot(online_pgc2[:, freq_band], color = 'r', ls='--')
ax[1].plot(residual[:, freq_band], color = 'm')
ax[1].set_ylabel('residual for band {:0.2f}'.format(processing.freqs[freq_band]))
for i in range(nloop):
ax[1].axvline(i*chunksize)
plt.show()
# TODO make one values per band
assert np.max(residual)<2e-2, 'hpaf online differt from offline'
import hearinglosssimulator as hls
import numpy as np
sample_rate = 44100.
duration = 4. #s
length = int(sample_rate * duration)
nb_channel = 2
# there are some helper for creating in_sounds in hls.in_soundgenerators
in_sound = hls.whitenoise(
length,
sample_rate=sample_rate,
)
#~ in_sound = hls.moving_erb_noise(length, trajectorytype='sinus', speed = .1)
# the shape must (length, nb_channel) so
in_sound = np.tile(in_sound[:, None], (1, nb_channel))
# define loss parameters
loss_params = {
'left': {
'freqs': [125., 250., 500., 1000., 2000., 4000., 8000.],
'compression_degree': [0., 0., 0., 0., 0., 0., 0.],
'passive_loss_db': [0., 0., 0., 0., 0., 0., 0.],
},
'right': {
'freqs': [125., 250., 500., 1000., 2000., 4000., 8000.],
'compression_degree': [0., 0., 0., 0., 0., 0., 0.],
'passive_loss_db': [0., 0., 0., 0., 0., 0., 0.],
}
}
def plot_residual():
nb_channel = 1
sample_rate = 44100.
chunksize = 256
#~ chunksize = 512
#~ chunksize = 1024
#~ chunksize =
#~ nloop = 200
nloop = 200
nb_freq_band = 10
length = int(chunksize * nloop)
in_buffer = hls.whitenoise(
length,
sample_rate=sample_rate,
)
in_buffer = np.tile(in_buffer[:, None], (1, nb_channel))
#~ lost_chunksize = np.linspace(0,1024, 5).astype(int)
lost_chunksize = np.arange(7).astype(int) * chunksize
#~ backward_chunksizes = [512,1024,1536,2048]
#~ backward_chunksizes = [1024,1536,2048]
#~ backward_chunksizes = np.linspace(1024,2048, 5).astype(int)
backward_chunksizes = lost_chunksize + chunksize
all_mean_residuals = np.zeros((len(backward_chunksizes), nb_freq_band))
all_max_residuals = np.zeros((len(backward_chunksizes), nb_freq_band))
for i, backward_chunksize in enumerate(backward_chunksizes):
print('backward_chunksize', backward_chunksize)
loss_params = {
'left': {
'freqs': [125., 250., 500., 1000., 2000., 4000., 8000.],
'compression_degree': [0., 0., 0., 0., 0., 0., 0.],
'passive_loss_db': [0., 0., 0., 0., 0., 0., 0.],
}
}
processing_conf = dict(nb_freq_band=nb_freq_band,
low_freq=40.,
high_freq=500.,
level_max=100,
level_step=100,
debug_mode=True,
chunksize=chunksize,
backward_chunksize=backward_chunksize,
loss_params=loss_params)
processing = hls.InvCGC(nb_channel=nb_channel,
sample_rate=sample_rate,
dtype='float32',
**processing_conf)
online_arrs = hls.run_instance_offline(
processing,
in_buffer,
chunksize,
sample_rate,
dtype='float32',
buffersize_margin=backward_chunksize)
#~ processing, online_arrs = hls.run_one_class_offline(hls.InvCGC, in_buffer, chunksize, sample_rate, processing_conf=processing_conf, buffersize_margin=backward_chunksize)
#~ freq_band = 2
online_hpaf = online_arrs['hpaf']
online_pgc2 = online_arrs['pgc2']
offline_pgc2 = online_pgc2.copy()
n = processing.nb_freq_band
for b in range(n):
offline_pgc2[:, b] = scipy.signal.sosfilt(
processing.coefficients_pgc[b, :, :], online_hpaf[::-1,
b])[::-1]
online_pgc2 = online_pgc2[:-backward_chunksize]
offline_pgc2 = offline_pgc2[:-backward_chunksize]
residual = np.abs(
(online_pgc2.astype('float64') - offline_pgc2.astype('float64')) /
np.mean(np.abs(offline_pgc2.astype('float64')), axis=0))
all_mean_residuals[i, :] = np.mean(residual, axis=0)
all_max_residuals[i, :] = np.max(residual, axis=0)
def my_imshow(m, ax):
im = ax.imshow(m,
interpolation='nearest',
origin='lower',
aspect='auto',
cmap='viridis') #, extent = extent, cmap=cmap)
im.set_clim(0, 0.05)
ax.set_xticks(np.arange(processing.freqs.size))
ax.set_xticklabels(['{:0.0f}'.format(f) for f in processing.freqs])
ax.set_yticks(np.arange(len(backward_chunksizes)))
ax.set_yticklabels(['{}'.format(f) for f in lost_chunksize])
ax.set_xlabel('freq')
ax.set_ylabel('lost_chunksize')
return im
print(all_max_residuals)
fig, axs = plt.subplots(nrows=2, sharex=True)
im1 = my_imshow(all_mean_residuals, axs[0])
im2 = my_imshow(all_max_residuals, axs[1])
cax = fig.add_axes([0.92, 0.05, .02, 0.9])
fig.colorbar(im1, ax=axs[0], cax=cax, orientation='vertical')
plt.show()