def TakeBF(self, time, show=False, saveIm=False):
""" Call signature example: BF = M.TakeBF(70)
-----------------------------------------
Return the beat frequencia of signal to the given instant (time).
Also is considered a mean of power spectrum of ten sweeps, to
avoid some resolution problems. """
S_mean_K = 0.;
S_mean_KA = 0.;
i1 = self.NearSweep(time);
i2 = i1 + int(self.SD.st * 1E-6 * self.SD.rate);
# time elapsed to next sweep in milliseconds
T_elapse = (self.SD.st + self.SD.si) * 1E-3;
# mean of 10 spectrograms.
for j in range(10):
# K band
S, f, t = mlab.specgram(self.SD.K[i1:i2], NFFT=self.nfft,
Fs=self.SD.rate, noverlap=self.nfft-self.fft_step,
pad_to=2**12);
S_mean_K += S / 10;
# Ka band
S, f, t = mlab.specgram(self.SD.KA[i1:i2], NFFT=self.nfft,
Fs=self.SD.rate, noverlap=self.nfft-self.fft_step,
pad_to=2**12);
S_mean_KA += S / 10;
# update indexes to next sweep.
i1 = self.NearSweep(time + T_elapse);
i2 = i1 + int(self.SD.st * 1E-6 * self.SD.rate);
T_elapse += (self.SD.st + self.SD.si) * 1E-3;
# avoid useles frequency depends on the case
if time < 10:
limSup = np.where(f > 1.65E7)[0].min();
limInf = np.where(f < 0.65E7)[0].max();
else:
limSup = np.where(f > 1.55E7)[0].min();
limInf = np.where(f < 0.35E7)[0].max();
# Take max line in spectrum.
freqIndexK = S_mean_K[limInf:limSup].argmax(axis=0) + limInf;
freqIndexKA = S_mean_KA[limInf:limSup].argmax(axis=0) + limInf;
# return to default image inferior limite
limSup = np.where(f > 1.55E7)[0].min();
limInf = np.where(f < 0.35E7)[0].max();
# Take frequency of max line.
BF_K = f[freqIndexK];
BF_KA = f[freqIndexKA];
# remove overlap if it has.
BF = np.concatenate([BF_K, BF_KA]);
if (saveIm or show):
S_K = S_mean_K[limInf:limSup];
S_KA = S_mean_KA[limInf:limSup];
f1 = f[limInf]; # Inferior figure limit
f2 = f[limSup]; # superior figure limit
self.__DrawImage(S_K, S_KA, time, f1, f2, BF, show, saveIm);
return BF;
def test_specgram(self):
for y, fstims in zip(self.y, self.fstimsall):
Pxx1, freqs1, t1 = mlab.specgram(y, NFFT=self.NFFT,
Fs=self.Fs,
noverlap=self.noverlap,
pad_to=self.pad_to,
sides='default')
Pxx1m = np.mean(Pxx1, axis=1)
np.testing.assert_array_equal(freqs1, self.freqss)
np.testing.assert_array_equal(t1, self.t)
# since we are using a single freq, all time slices should be
# about the same
np.testing.assert_allclose(np.diff(Pxx1, axis=1).max(), 0,
atol=1e-08)
for fstim in fstims:
i = np.abs(freqs1 - fstim).argmin()
self.assertTrue(Pxx1m[i] > Pxx1m[i+1])
self.assertTrue(Pxx1m[i] > Pxx1m[i-1])
Pxx2, freqs2, t2 = mlab.specgram(y, NFFT=self.NFFT,
Fs=self.Fs,
noverlap=self.noverlap,
pad_to=self.pad_to,
sides='onesided')
Pxx2m = np.mean(Pxx2, axis=1)
np.testing.assert_array_equal(freqs2, self.freqss)
np.testing.assert_array_equal(t2, self.t)
np.testing.assert_allclose(np.diff(Pxx2, axis=1).max(), 0,
atol=1e-08)
for fstim in fstims:
i = np.abs(freqs2 - fstim).argmin()
self.assertTrue(Pxx2m[i] > Pxx2m[i+1])
self.assertTrue(Pxx2m[i] > Pxx2m[i-1])
Pxx3, freqs3, t3 = mlab.specgram(y, NFFT=self.NFFT,
Fs=self.Fs,
noverlap=self.noverlap,
pad_to=self.pad_to,
sides='twosided')
Pxx3m = np.mean(Pxx3, axis=1)
np.testing.assert_array_equal(freqs3, self.freqsd)
np.testing.assert_array_equal(t3, self.t)
np.testing.assert_allclose(np.diff(Pxx3, axis=1).max(), 0,
atol=1e-08)
for fstim in fstims:
i = np.abs(freqs3 - fstim).argmin()
self.assertTrue(Pxx3m[i] > Pxx3m[i+1])
self.assertTrue(Pxx3m[i] > Pxx3m[i-1])
def fingerprint(channel_samples, Fs=DEFAULT_FS, wsize=DEFAULT_WINDOW_SIZE, wratio=DEFAULT_OVERLAP_RATIO, fan_value=DEFAULT_FAN_VALUE, amp_min=DEFAULT_AMP_MIN):
"""
FFT the channel, loklsg transform output, find local maxima, then return
locally sensitive hashes.
"""
# FFT the signal and extract frequency components
arr2D = mlab.specgram(
channel_samples,
NFFT=wsize,
Fs=Fs,
window=mlab.window_hanning,
noverlap=int(wsize * wratio))[0]
# apply log transform since specgram() returns linear array
np.seterr(all='ignore')
arr2D = 10 * np.log10(arr2D)
arr2D[arr2D == -np.inf] = 0 # replace infs with zeros
# find local maxima
local_maxima = get_2D_peaks(arr2D, plot=False, amp_min=amp_min)
# return hashes
return generate_hashes(local_maxima, fan_value=fan_value)
def get_fingerprints(X,
ws=WINDOW_SIZE,
sr=SAMPLE_RATE,
si=SLIDE_INTERVAL,
ns=NEIGHBOR_SIZE,
thres=MIN_AMPLITUDE,
fo=FAN_OUT):
"""
Calculate the fingerprints (hash values) of a given signal X.
Return the fingerprints in term of (sha256, diff time).
"""
# apply fft to X and retrieve the spectrogram of X
sg, freqs, t = specgram(X,
NFFT=ws,
Fs=sr,
noverlap=si)
# apply log transform to the spectrogram
sg = 10 * log10(sg)
# find the local maxima in neighborhood
peaks = find_peaks(sg, ns, thres)
# generate final fingerprints in form of (sha256, diff time)
fingerprints = generate_fingerprints(peaks, fo)
return fingerprints
def transform_one(row):
row -= MIN
row /= MAX
spec = mlab.specgram(row, NFFT=256, Fs=16384)[0]
spec = np.log(spec)#.ravel()
#spec = ndimage.gaussian_filter(spec, 1)
return spec
def simple_spectogram(df, meta):
import numpy as np
import pandas as pd
from matplotlib.mlab import specgram
def merge_two_dicts(x, y):
z = x.copy()
z.update(y)
return z
# extract numpy array from pandas.DataFrame
data = np.asarray(df)
spectrum, freqs, t = specgram(data.squeeze())
# Return value must be of a sequence of pandas.DataFrame
df = pd.DataFrame(spectrum)
# updata meta
new_meta = {'spectogram_out': {'frequencies': freqs.tolist(),
'time_points': t.tolist()}}
d = merge_two_dicts(meta.to_dict(), new_meta)
meta = pd.DataFrame(d)
return df, meta
def _energy_func(self, x, **kwargs):
from matplotlib.mlab import specgram
rval = sp.zeros_like(x)
ns, nc = x.shape
ov_samples = 0
offset = 0
if self.overlap == 1:
ov_samples = self.nfft * 0.5
offset = self.nfft / 4
elif self.overlap == 2:
ov_samples = self.nfft * 0.75
offset = self.nfft * 0.375
step = self.nfft - ov_samples
for c in xrange(nc):
psd_arr, freqs, times = specgram(x[:, c], NFFT=self.nfft, Fs=self.srate, noverlap=ov_samples)
mask = freqs < self.cutoff_hz
for b in xrange(len(times)):
bin_s = b * step + offset
bin_e = bin_s + step
if self.en_func == 'mean_coeff':
rval[bin_s:bin_e, c] = psd_arr[mask == True, b].mean() / psd_arr[mask == False, b].mean()
elif self.en_func == 'max_coeff':
rval[bin_s:bin_e, c] = psd_arr[mask == True, b].max() / psd_arr[mask == False, b].max()
elif self.en_func == 'max_normed':
rval[bin_s:bin_e, c] = psd_arr[mask == True, b].max() / psd_arr[:, b].sum(axis = 0)
else:
raise RuntimeError('Energy function does not exist!')
return rval
def psd_eeg(data, **kwargs):
for arg in kwargs:
if arg == "NFFT":
NFFT = int(kwargs[arg])
if arg == "dt":
Fs = 1.0 / float(kwargs[arg])
if arg == "noverlap":
noverlap = int(kwargs[arg])
px_list = []
print "PSD Computing..."
for sample in range(data.shape[0]):
for ch in range(data.shape[1]):
eeg = data[sample, ch, :]
[Pxx, freq, t] = specgram(eeg, NFFT=NFFT, noverlap=noverlap, Fs=Fs)
px_list.append(Pxx)
shape = px_list[0].shape
pot = np.array(px_list)
pot = pot.reshape(data.shape[0], data.shape[1], shape[0], -1)
del px_list
return [pot, freq]
def _do_fft(self):
t = self.getp('t')
v = self.getp('v')
try:
Fs = 1/(t[1] - t[0])
NFFT = self.getp('NFFT')
detrend = detrends[self.getp('detrend')]
window = windows[self.getp('window')]
noverlap = self.getp('noverlap')
sides = self.getp('sides')
pad_to = self.getp('pad_to')
if pad_to == 'none': pad_to = None
z, y, x = mlab.specgram(v, NFFT=NFFT,
Fs=Fs,
detrend = detrend,
window = window,
noverlap = noverlap,
pad_to=pad_to,
sides = sides,
scale_by_freq=None)
self.setp("x", x + t[0])
self.setp("y", y)
self.setp("z", z)
except:
traceback.print_exc()
self.setp("x", None)
self.setp("y", None)
self.setp("z", None)
return False
return True
def calculate_specgram(self, nfft, noverlap, **kwargs):
spec = mlab.specgram(self.audio, NFFT=nfft, Fs=self.framerate, noverlap=noverlap, **kwargs)
self.specgram_nfft = nfft
self.specgram_overlap = noverlap
self.specgram = spec[0]
self.specgram_freqs = spec[1]
self.specgram_bins = spec[2]
def fingerprint(
channel_samples,
Fs=DEFAULT_FS,
wsize=DEFAULT_WINDOW_SIZE,
wratio=DEFAULT_OVERLAP_RATIO,
fan_value=DEFAULT_FAN_VALUE,
amp_min=DEFAULT_AMP_MIN,
plot=False,
):
"""
FFT the channel, log transform output, find local maxima, then return
locally sensitive hashes.
"""
# FFT the signal and extract frequency components
# http://matplotlib.org/api/mlab_api.html#matplotlib.mlab.specgram
# return of specgram is (spectrum, freqs, t)
arr2D, freqs, times = mlab.specgram(
channel_samples, NFFT=wsize, Fs=Fs, window=mlab.window_hanning, noverlap=int(wsize * wratio)
)
# apply log transform since specgram() returns linear array
arr2D = 10 * np.log10(arr2D)
arr2D[arr2D == -np.inf] = 0 # replace infs with zeros
# find local maxima
local_maxima = get_2D_peaks(arr2D, plot=plot, amp_min=amp_min, freqs=freqs, times=times)
# return hashes
return generate_hashes(local_maxima, fan_value=fan_value)
def plot_specgram(ax, data, fs, nfft=256, noverlap=128, window='hann',
cmap='jet', interpolation='bilinear', rasterized=True):
if window not in SPECGRAM_WINDOWS:
raise ValueError("Window not supported")
elif window == "boxcar":
mwindow = signal.boxcar(nfft)
elif window == "hamming":
mwindow = signal.hamming(nfft)
elif window == "hann":
mwindow = signal.hann(nfft)
elif window == "bartlett":
mwindow = signal.bartlett(nfft)
elif window == "blackman":
mwindow = signal.blackman(nfft)
elif window == "blackmanharris":
mwindow = signal.blackmanharris(nfft)
specgram, freqs, time = mlab.specgram(data, NFFT=nfft, Fs=fs,
window=mwindow,
noverlap=noverlap)
specgram = 10 * np.log10(specgram[1:, :])
specgram = np.flipud(specgram)
freqs = freqs[1:]
halfbin_time = (time[1] - time[0]) / 2.0
halfbin_freq = (freqs[1] - freqs[0]) / 2.0
extent = (time[0] - halfbin_time, time[-1] + halfbin_time,
freqs[0] - halfbin_freq, freqs[-1] + halfbin_freq)
ax.imshow(specgram, cmap=cmap, interpolation=interpolation,
extent=extent, rasterized=rasterized)
ax.axis('tight')
def load_spec(path): ''' Loads the spectrogram file that was saved by the function sess_spectrogram Input: path : either the absolute path for the spectrogram file or a list containing the cage ID, block number, and sess Output: P : MxN spectrogram array (M frequencies, N time bins) F : vector of length M indicating the frequencies included in the spectrogram ''' exten = os.path.splitext(path)[-1] if exten == '.npz': tmp = np.load(path) P = np.array(tmp['P'].min().todense()) F = tmp['F'] T = tmp['T'] tmp.close() elif exten == '.h5': NFFT = 512 noverlap = 256 f = h5py.File(path, 'r+') s = f['s'].value fs = f['fs'].value f.close() (P, F, T) = specgram(s, Fs = fs, NFFT = NFFT, noverlap = noverlap) return P, F, T
def get_fingerprint_from_data(wave_data):
# pxx[freq_idx][t] - мощность сигнала
pxx, _, _ = mlab.specgram(
wave_data,
NFFT=cf.WINDOW_SIZE,
noverlap=cf.WINDOW_OVERLAP,
Fs=cf.SAMPLE_RATE)
# 300-2870 | delta = 256 * 10 = 8 * 32 * 10
matrix = pxx[30*2:300*2].transpose()
cnt1, cnt2 = 0, 0
arr = []
for time, timeline in enumerate(matrix):
if time == 0:
continue
hash64, pow2 = 0, 1
for j in range(1, 65):
energy1 = energy(matrix, time, j) - energy(matrix, time, j - 1)
energy2 = energy(matrix, time - 1, j) - energy(matrix, time - 1, j - 1)
if energy1 - energy2 > 0:
hash64 += pow2
cnt1 += 1
else:
cnt2 += 1
pow2 *= 2
arr.append(hash64)
print('Done fingerprinting...', cnt1, cnt2)
return [arr]
def _plot_spectrum(self, data, sampling_frequency):
(Pxx, freqs, bins) = mlab.specgram(data,
Fs=sampling_frequency,
NFFT=self.NFFT,
noverlap=self.noverlap)
if numpy.any(Pxx[0,0] == 0):
self._log("SpectrumPlot::Instance has power 0 in a frequency band, skipping...")
return
# Update minimal and maximal value
self.min_value = min(self.min_value, min(Pxx.flatten()))
self.max_value = max(self.max_value, max(Pxx.flatten()))
Z = numpy.flipud(Pxx)
extent = 0, numpy.amax(bins), freqs[0], freqs[-1]
pylab.imshow(Z, None, extent=extent, vmin=self.min_value,
vmax=self.max_value)
pylab.axis('auto')
pylab.xlabel("Time(s)")
pylab.ylabel("Frequency(Hz)")
if self.colorbar and not self.physiological_arrangement:
pylab.colorbar()
return (Pxx, freqs, bins)
def calc_power_spectrum(lfp): npts, ntrials = lfp.shape Fs = 384. X = np.zeros() x = np.zeros((129, 7, ntrials)) for i in range(ntrials): x[:, :, i] = specgram(lfp[:127, i], Fs = Fs)[0]
def specgram(signal, sampling_frequency, time_resolution,
frequency_resolution, bath_signals=[],
high_frequency_cutoff=None, axes=None, logscale=True, **kwargs):
"""
This function wraps matplotlib.mlab.specgram to provide a more intuitive
interface.
Inputs:
signal : the input signal (a one dimensional array)
sampling_frequency : the sampling frequency of signal
time_resolution : the desired time resolution of the specgram
this is the guaranteed worst time resolution
frequency_resolution : the desired frequency resolution of the
specgram. this is the guaranteed worst
frequency resolution.
--keyword arguments--
bath_signals : Subtracts a bath signal from the spectrogram
axes=None : If an Axes instance is passed then it will
plot to that.
**kwargs : Arguments passed on to
matplotlib.mlab.specgram
Returns:
If axes is None:
Pxx
freqs
bins
if axes is an Axes instance:
Pxx, freqs, bins, and im
"""
if (high_frequency_cutoff is not None
and high_frequency_cutoff < sampling_frequency):
resampled_signal = resample_signal(signal, sampling_frequency,
high_frequency_cutoff)
else:
high_frequency_cutoff = sampling_frequency
resampled_signal = signal
num_data_samples = len(resampled_signal)
specgram_settings = find_NFFT_and_noverlap(frequency_resolution,
high_frequency_cutoff,
time_resolution,
num_data_samples)
NFFT = specgram_settings['power_of_two_NFFT']
noverlap = specgram_settings['noverlap']
Pxx, freqs, bins = mlab.specgram(resampled_signal,
NFFT=NFFT,
Fs=high_frequency_cutoff,
noverlap=noverlap, **kwargs)
plotted_Pxx = Pxx
if bath_signals:
bath_signal = numpy.hstack(bath_signals)
psd_Pxx, psd_freqs = psd(bath_signal, sampling_frequency,
frequency_resolution,
high_frequency_cutoff=high_frequency_cutoff )
plotted_Pxx = (Pxx.T/psd_Pxx).T
if axes is not None:
im = plot_specgram(plotted_Pxx, freqs, bins, axes, logscale=logscale)
return plotted_Pxx, freqs, bins, im
return plotted_Pxx, freqs, bins
def score(self, audio):
nfft = int(self.window*audio.framerate)
audio.calculate_specgram(nfft=nfft, noverlap=nfft/2)
freqs = np.where((audio.specgram_freqs >= self.lower_call_frequency)*(audio.specgram_freqs <= self.upper_call_frequency))
spec2 = mlab.specgram(mean(log(audio.specgram[freqs[0],]), 0), NFFT=1024, noverlap=512, Fs=2/self.window)
freqs2 = np.where((spec2[1] >= self.lower_syllable_frequency)*(spec2[1] <= self.upper_syllable_frequency))
max_kiwi = max(np.max(spec2[0][freqs2[0], :], 0))
mean_kiwi = np.exp(np.mean(np.mean(np.log(spec2[0][freqs2[0], :]), 0)))
return max_kiwi/mean_kiwi
def generateSpectogram(channel_samples, Fs=DEFAULT_FS, wsize=DEFAULT_WINDOW_SIZE, wratio=DEFAULT_OVERLAP_RATIO, fan_value=DEFAULT_FAN_VALUE, amp_min=DEFAULT_AMP_MIN): # a tuple (Pxx, freqs, t) is said to be the return value of specgram arr2D = mlab.specgram(channel_samples, NFFT=wsize, Fs=Fs, window=mlab.window_hanning, noverlap=int(wsize * wratio))[0] # apply log transform since specgram() returns linear array arr2D = 10 * np.log10(arr2D) arr2D[arr2D == -np.inf] = 0 # replace infs with zeros # find local maxima plotPeaks(arr2D, amp_min=amp_min)
def __call__(iterator):
for AudioFile in iterator:
x = AudioFile.data
data = mlab.specgram(x, NFFT=256, Fs=2, detrend=mlab.detrend_none,
window=mlab.window_hanning, noverlap=128,
cmap=None, xextent=None, pad_to=None, sides='default',
scale_by_freq=None, mode='default')
yield data
def get_spectrum_data(self):
print("Calculating spectrum data...")
self.spec_PSDperHz, self.spec_freqs, self.spec_t = mlab.specgram(np.squeeze(self.data),
NFFT=self.NFFT,
window=mlab.window_hanning,
Fs=self.fs_Hz,
noverlap=self.overlap
) # returns PSD power per Hz
# convert the units of the spectral data
self.spec_PSDperBin = self.spec_PSDperHz * self.fs_Hz / float(self.NFFT)
def get_specgram(self):
winpoints = self.winfunc(self.specfftsize)
iq_spec, f, t = mlab.specgram(self.iq, Fs=self.sample_rate,
NFFT=self.specfftsize,
noverlap=self.specfftsize/4.0,
window=winpoints,
scale_by_freq=False)
self.iq_spec = 10.0*scipy.log10(abs(iq_spec))
self.spec_f = f
self.spec_t = t
def specgram(
signal,
sampling_frequency,
time_resolution,
frequency_resolution,
high_frequency_cutoff=None,
logscale=True,
**kwargs
):
"""This function wraps matplotlib.mlab.psd to provide a more intuitive
interface.
Plot with:
power, freqs, bins = specgram(...)
extent = (bins[0], bins[-1], freqs[0], freqs[-1])
imshow(power, aspect='auto', origin='lower', extent=extent) # from pyplot
:param: signal - the input signal (a one dimensional array)
:param: sampling_frequency - the sampling frequency of signal (i.e.: 10000)
:param: frequency_resolution - the desired frequency resolution of the specgram.
this is the guaranteed worst frequency resolution.
:param: time_resolution - the desired frequency resolution of the specgram.
this is the guaranteed worst time resolution.
:param: high_frequency_cutoff - optional high freq. cutoff. resamples data
to this value and then uses that for Fs parameter
:param: logscale - rescale data based on log values? defaults is True
:param: **kwargs - Arguments passed on to matplotlib.mlab.psd
:returns: - tuple of three numpy arrays:
power - 2d array of power (dB/Hz)
freqs - in Hz
bins - in seconds
"""
if high_frequency_cutoff is not None and high_frequency_cutoff < sampling_frequency:
resampled_signal = resample_signal(signal, sampling_frequency, high_frequency_cutoff)
else:
high_frequency_cutoff = sampling_frequency
resampled_signal = signal
num_data_samples = len(resampled_signal)
specgram_settings = find_NFFT_and_noverlap(
frequency_resolution, high_frequency_cutoff, time_resolution, num_data_samples
)
NFFT = specgram_settings["power_of_two_NFFT"]
noverlap = specgram_settings["noverlap"]
power, freqs, bins = mlab.specgram(
resampled_signal, NFFT=NFFT, Fs=high_frequency_cutoff, noverlap=noverlap, **kwargs
)
if logscale:
power = 10 * np.log10(power)
return power, freqs, bins
def ReadAndAnalyze(f):
wav = wave.open(f, 'r')
frames = wav.readframes(-1)
sound_info = numpy.fromstring(frames, 'Int16')
frame_rate = wav.getframerate()
wav.close()
specdata = mlab.specgram(sound_info,
NFFT=nfft,
pad_to=padto,
Fs=frame_rate)
spectrum = specdata[0]
freqs = specdata[1]
return spectrum, freqs
def get_center_frequencies(self):
"""
Returns center frequencies of filter bank channels
"""
nwin, nshift, nfft, fs, noverlap = self.__get_fft_params()
_, fc, _ = specgram(np.zeros((2*nwin,)), NFFT=nfft, Fs=fs,
noverlap=noverlap)
if self.f_cutoff is not None:
valid = np.logical_and(fc >= self.f_cutoff[0],
fc <= self.f_cutoff[1])
fc = fc[valid]
return fc
def get_power_spectrum_of_wave(sample_rate, data):
noverlap = 96
nfft = 128
spectrum, frequencies, times = specgram(
data,
Fs=sample_rate,
noverlap=noverlap,
NFFT=nfft,
scale_by_freq=True,
)
return spectrum
def transform(self, waveform):
"""Converts a waveform to a suitable spectrogram.
Removes high and low frequencies, rebins in time (via median)
to reduce data size. Returned times are the midpoints of the new bins.
Returns: Pxx, freqs, t
Pxx is an array of dB power of the shape (len(freqs), len(t)).
It will be real but may contain -infs due to log10
"""
# For now use NFFT of 256 to get appropriately wide freq bands, then
# downsample in time
Pxx, freqs, t = mlab.specgram(waveform, NFFT=self.NFFT,
noverlap=self.noverlap, Fs=self.Fs, detrend=self.detrend,
**self.specgram_kwargs)
# Apply the normalization
Pxx = Pxx * np.tile(freqs[:, np.newaxis] ** self.normalization,
(1, Pxx.shape[1]))
# strip out unused frequencies
if self.max_freq is not None:
Pxx = Pxx[freqs < self.max_freq, :]
freqs = freqs[freqs < self.max_freq]
if self.min_freq is not None:
Pxx = Pxx[freqs > self.min_freq, :]
freqs = freqs[freqs > self.min_freq]
# Rebin in size "downsample_ratio". If last bin is not full, discard.
Pxx_rebinned = []
t_rebinned = []
for n in range(0, len(t) - self.downsample_ratio + 1,
self.downsample_ratio):
Pxx_rebinned.append(
np.median(Pxx[:, n:n+self.downsample_ratio], axis=1).flatten())
t_rebinned.append(
np.mean(t[n:n+self.downsample_ratio]))
# Convert to arrays
Pxx_rebinned_a = np.transpose(np.array(Pxx_rebinned))
t_rebinned_a = np.array(t_rebinned)
# log it and deal with infs
Pxx_rebinned_a_log = -np.inf * np.ones_like(Pxx_rebinned_a)
Pxx_rebinned_a_log[np.nonzero(Pxx_rebinned_a)] = \
10 * np.log10(Pxx_rebinned_a[np.nonzero(Pxx_rebinned_a)])
self.freqs = freqs
self.t = t_rebinned_a
return Pxx_rebinned_a_log, freqs, t_rebinned_a
def __init_spectrogram(self,
nfft = DEFAULT_WINDOW_SIZE,
overlap = DEFAULT_OVERLAP_RATIO):
# matplotlib has 3 different specgram implementations - mlab, pyplot, and on
# the axes:
# 1. mlab - [https://github.com/matplotlib/matplotlib/blob/78f1942f5af063abacc3d5709912bde6bbb6ffec/lib/matplotlib/mlab.py]
# 2. pyplot - [https://github.com/matplotlib/matplotlib/blob/e7717f71abae542ba0a5fea5b072bf52548d5ee3/lib/matplotlib/pyplot.py]
# 3. axes - [https://github.com/matplotlib/matplotlib/blob/e7717f71abae542ba0a5fea5b072bf52548d5ee3/lib/matplotlib/axes/_axes.py]
# 1 is the primary implementation. 2 calls 3 and 3 calls 1, but does the
# added job of plotting after log scaling, and flipping the surface
# matrix appropriately. We derive the plotting code from theirs.
Pxx, freqs, bins = mlab.specgram(self.data, NFFT = nfft, Fs = self.frequency, pad_to = self.specgram_freq_bins, noverlap = (int(nfft*overlap)))
return Pxx, freqs, bins
def trial_specgram(d, samplerate=None, NFFT=256):
'''
Calculate a spectrogram for each channel and each trial.
Parameters
----------
d : :class:`DataSet`
The trials.
samplerate : float
The sample rate of the data. When omitted,
:func:`psychic.get_samplerate` is used to estimate the sample rate.
NFFT : int
Number of FFT points to use to calculate the spectrograms.
Returns
-------
d : :class:`DataSet`
The spectrograms:
- ``d.data``: [channels x freqs x samples x trials]
- ``d.feat_lab``: The feature labels for the axes [channels
(strings), frequencies (floats), time in seconds (floats)]
'''
assert d.data.ndim == 3
if samplerate is None:
assert d.feat_lab is not None, 'Must either supply samplerate or feat_lab to deduce it.'
samplerate = np.round(1./np.median(np.diff([float(x) for x in d.feat_lab[1]])))
all_TFs = []
for trial in range(d.ninstances):
channel_TFs = []
for channel in range(d.data.shape[0]):
TF, freqs, times = specgram(d.data[channel,:,trial], NFFT, samplerate, noverlap=NFFT/2)
channel_TFs.append(TF.T[np.newaxis,:,:])
all_TFs.append(np.concatenate(channel_TFs, axis=0)[..., np.newaxis])
all_TFs = np.concatenate(all_TFs, axis=3)
nchannels, nfreqs, nsamples, ninstances = all_TFs.shape
feat_lab = [d.feat_lab[0], times.tolist(), freqs.tolist()]
feat_dim_lab=['channels', 'time', 'frequencies']
return DataSet(
data=all_TFs,
feat_lab=feat_lab,
feat_dim_lab=feat_dim_lab,
default=d)
def showSignal(self):
Pxx, freqs, bins = specgram(self.signal, NFFT = self.NFFT, Fs= self.fs, noverlap = 2*self.NFFT/3, window = gaussian(self.NFFT,self.NFFT/6))
Z= 10*np.log10(Pxx)
Z = np.flipud(Z)
Z[Z>np.amax(Z)-3]= np.amax(Z)-3
self.rawSpecData = Z
Z[Z<(np.amax(Z)-self.dynamicRange)]=self.noiseFloor
extent = 0, self.timeLength, freqs[0], freqs[-1]
self.axes.imshow(Z,cm.get_cmap('Greys'),extent = extent)
self.axes.axis('auto')
self.axes.set_ylim([0,self.freqRange])
self.axes.set_xlim([0,self.timeLength])
self.axes.grid(True)
self.draw()
def spectrogram(data,
samp_rate,
per_lap=0.9,
wlen=None,
log=False,
outfile=None,
fmt=None,
axes=None,
dbscale=False,
mult=8.0,
cmap=obspy_sequential,
zorder=None,
title=None,
show=True,
clip=[0.0, 1.0]):
"""
Computes and plots spectrogram of the input data.
:param data: Input data
:type samp_rate: float
:param samp_rate: Samplerate in Hz
:type per_lap: float
:param per_lap: Percentage of overlap of sliding window, ranging from 0
to 1. High overlaps take a long time to compute.
:type wlen: int or float
:param wlen: Window length for fft in seconds. If this parameter is too
small, the calculation will take forever. If None, it defaults to
(samp_rate/100.0).
:type log: bool
:param log: Logarithmic frequency axis if True, linear frequency axis
otherwise.
:type outfile: str
:param outfile: String for the filename of output file, if None
interactive plotting is activated.
:type fmt: str
:param fmt: Format of image to save
:type axes: :class:`matplotlib.axes.Axes`
:param axes: Plot into given axes, this deactivates the fmt and
outfile option.
:type dbscale: bool
:param dbscale: If True 10 * log10 of color values is taken, if False the
sqrt is taken.
:type mult: float
:param mult: Pad zeros to length mult * wlen. This will make the
spectrogram smoother.
:type cmap: :class:`matplotlib.colors.Colormap`
:param cmap: Specify a custom colormap instance. If not specified, then the
default ObsPy sequential colormap is used.
:type zorder: float
:param zorder: Specify the zorder of the plot. Only of importance if other
plots in the same axes are executed.
:type title: str
:param title: Set the plot title
:type show: bool
:param show: Do not call `plt.show()` at end of routine. That way, further
modifications can be done to the figure before showing it.
:type clip: [float, float]
:param clip: adjust colormap to clip at lower and/or upper end. The given
percentages of the amplitude range (linear or logarithmic depending
on option `dbscale`) are clipped.
"""
import matplotlib.pyplot as plt
# enforce float for samp_rate
samp_rate = float(samp_rate)
# set wlen from samp_rate if not specified otherwise
if not wlen:
wlen = samp_rate / 100.
npts = len(data)
# nfft needs to be an integer, otherwise a deprecation will be raised
# XXX add condition for too many windows => calculation takes for ever
nfft = int(_nearest_pow_2(wlen * samp_rate))
if nfft > npts:
nfft = int(_nearest_pow_2(npts / 8.0))
if mult is not None:
mult = int(_nearest_pow_2(mult))
mult = mult * nfft
nlap = int(nfft * float(per_lap))
data = data - data.mean()
end = npts / samp_rate
# Here we call not plt.specgram as this already produces a plot
# matplotlib.mlab.specgram should be faster as it computes only the
# arrays
# XXX mlab.specgram uses fft, would be better and faster use rfft
specgram, freq, time = mlab.specgram(data,
Fs=samp_rate,
NFFT=nfft,
pad_to=mult,
noverlap=nlap)
# db scale and remove zero/offset for amplitude
if dbscale:
specgram = 10 * np.log10(specgram[1:, :])
else:
specgram = np.sqrt(specgram[1:, :])
freq = freq[1:]
vmin, vmax = clip
if vmin < 0 or vmax > 1 or vmin >= vmax:
msg = "Invalid parameters for clip option."
raise ValueError(msg)
_range = float(specgram.max() - specgram.min())
vmin = specgram.min() + vmin * _range
vmax = specgram.min() + vmax * _range
norm = Normalize(vmin, vmax, clip=True)
if not axes:
fig = plt.figure()
ax = fig.add_subplot(111)
else:
ax = axes
# calculate half bin width
halfbin_time = (time[1] - time[0]) / 2.0
halfbin_freq = (freq[1] - freq[0]) / 2.0
# argument None is not allowed for kwargs on matplotlib python 3.3
kwargs = {
k: v
for k, v in (('cmap', cmap), ('zorder', zorder)) if v is not None
}
if log:
# pcolor expects one bin more at the right end
freq = np.concatenate((freq, [freq[-1] + 2 * halfbin_freq]))
time = np.concatenate((time, [time[-1] + 2 * halfbin_time]))
# center bin
time -= halfbin_time
freq -= halfbin_freq
# Log scaling for frequency values (y-axis)
ax.set_yscale('log')
# Plot times
ax.pcolormesh(time, freq, specgram, norm=norm, **kwargs)
else:
# this method is much much faster!
specgram = np.flipud(specgram)
# center bin
extent = (time[0] - halfbin_time, time[-1] + halfbin_time,
freq[0] - halfbin_freq, freq[-1] + halfbin_freq)
ax.imshow(specgram, interpolation="nearest", extent=extent, **kwargs)
# set correct way of axis, whitespace before and after with window
# length
ax.axis('tight')
ax.set_xlim(0, end)
ax.grid(False)
if axes:
return ax
ax.set_xlabel('Time [s]')
ax.set_ylabel('Frequency [Hz]')
if title:
ax.set_title(title)
if not os.environ.get('SPHINXBUILD'):
# ignoring all NumPy warnings during plot
with np.errstate(all='ignore'):
plt.draw()
if outfile:
if fmt:
fig.savefig(outfile, format=fmt)
else:
fig.savefig(outfile)
elif show:
plt.show()
else:
return fig
def audioToFilteredSpectrogram(data,
expandByOne=True,
dropZeroColumnsPercent=0.95):
# calculate the spectogram
#
#tempSpec = np.log10(mlab.specgram(data, NFFT=512, noverlap=256, Fs=16000)[0])
tempSpec = np.log10(
mlab.specgram(data, NFFT=512, noverlap=256, Fs=16000)[0] + 1)
# drop higher frequencies
tempSpec = tempSpec[0:200, :]
tempSpecFiltered = np.copy(tempSpec)
# we analize the spectogram by 20x30 sized cells
# to achieve better accuray the size of this cell should be finetuned
rowBorders = np.ceil(np.linspace(0, tempSpec.shape[0], 20))
columnBorders = np.hstack((np.ceil(np.arange(0, tempSpec.shape[1],
30)), tempSpec.shape[1]))
rowBorders = [int(x) for x in rowBorders]
columnBorders = [int(x) for x in columnBorders]
keepCells = np.ones((len(rowBorders) - 1, len(columnBorders) - 1))
# we create a mask for the spectogram: we scan the spectogram with the 20x30 sized
# cell and create 0 mask based on the mean and std of the spectogram calculated for the cells and rows
for i in range(len(rowBorders) - 1):
row_mean = np.mean(tempSpec[rowBorders[i]:rowBorders[i + 1], :])
row_std = np.std(tempSpec[rowBorders[i]:rowBorders[i + 1], :])
for j in range(len(columnBorders) - 1):
cell_mean = np.mean(tempSpec[rowBorders[i]:rowBorders[i + 1],
columnBorders[j]:columnBorders[j +
1]])
cell_max_top10_mean = np.mean(
np.sort(tempSpec[rowBorders[i]:rowBorders[i + 1],
columnBorders[j]:columnBorders[j + 1]],
axis=None)[-10:])
if (cell_mean < 0 or ((cell_max_top10_mean) <
(row_mean + row_std) * 1.5)):
keepCells[i, j] = 0
# expand by ones (see above)
if expandByOne:
keepCells = expandOnes(keepCells)
# apply the mask to the spectogram
for i in range(keepCells.shape[0]):
for j in range(keepCells.shape[1]):
if not keepCells[i, j]:
tempSpecFiltered[rowBorders[i]:rowBorders[i + 1],
columnBorders[j]:columnBorders[j + 1]] = 0
# drop zero columns
# the amount of zero values along axis 0 (frequency) is calculated for every column (timeslice)
# and it is dropped, if the number of zero values is higher than the dropZeroColumnsPercent
# eg. dropZeroColumnsPercent=0.95, than a column (timeslice) is dropped, if more than 95% of the values (frequencies) is 0
tempSpecFilteredBackup = np.copy(tempSpecFiltered)
tempSpecFiltered = np.delete(
tempSpecFiltered,
np.nonzero((tempSpecFiltered == 0).sum(
axis=0) > tempSpecFiltered.shape[0] * dropZeroColumnsPercent),
axis=1)
# if every row was 0 than use the backed up spectogram
if tempSpecFiltered.shape[1] == 0:
tempSpecFiltered = tempSpecFilteredBackup
return tempSpec, tempSpecFiltered
def spectrogram(x, SR, winlen):
Pxx, freqs, bins, im = mlab.specgram(x, NFFT=winlen, Fs=SR)
def features_and_labels(path_input, path_label, access, type_data):
#This function computes spectrogram for each audio, creates the extended featureMAP, divides this in segments of lenght 400
#and creates the list of the inputs of the neural network and the list of the rispective labels
#The lists are saved of hdf5 files
#type_data defines the partition (train, dev, eval)
list_input=[]
list_label_b=[]
list_label_m=[]
UFeatureMAP=np.zeros((257, 1600))
segment=np.zeros((257,400))
M=400 #Lenght of the segments
L=200 #Overlap between segments
file_label = open(path_label,"r", encoding='utf8')
rows=file_label.readlines();
file_label.close()
#File txt where for each audio where are saved the number of segments and the label
file_segments=type_data+'_segments.txt'
f=open(file_segments,'w', encoding='utf8')
f.close()
#File where segments and labels will be saved
filename_input=os.path.join(type_data+'_input.hdf5') #es. train_input.hdf5
filename_label_m=os.path.join(type_data+'_label_m.hdf5')
filename_label_b=os.path.join(type_data+'_label_b.hdf5')
fx = h5py.File(filename_input, "a")
fy_multi = h5py.File(filename_label_m, "a")
fy_binary = h5py.File(filename_label_b, "a")
dset = fx.create_dataset('dataset', (0,257,400,1), maxshape=(None,None,None,None), chunks=True,
compression="gzip", compression_opts=9)
dset_binary=fy_binary.create_dataset('dataset', (0,), maxshape=(None,), chunks=True,
compression="gzip", compression_opts=4)
dset_multi=fy_multi.create_dataset('dataset', (0,), maxshape=(None,), chunks=True,
compression="gzip", compression_opts=4)
print('Start extracting features...')
print('Creating hdf5 file...')
N_files=len(glob.glob(path_input))
S=3000 #How much segments are appended to the file hdf5 each time
N_iterazioni=int(N_files/S)
rest2=N_files-N_iterazioni*S
counter=0
#glob.glob return a string like this: /content/LA/ASVspoof2019_LA_train/flac/LA_T_1929428.flac
#We wanto to extract from this string the name of the audio that is in this case: LA_T_1929428
#So, to do that we take filename[len(filename)-17:-5]
for filename in glob.glob(path_input):
counter=counter+1
flag_found=0
for row in rows:
fields=row.split(" ")
if fields[1]==filename[len(filename)-17:-5]:
flag_found=1
label_multi=fields[3]
label_multi=convert_label(label_multi, access)
if fields[4]=='bonafide\n':
label_binary=0
if fields[4]=='spoof\n':
label_binary=1
break
if flag_found==1:
#Spectrogram
data,fs=sf.read(filename)
[Ps, f, t] = mlab.specgram(data, NFFT=512, Fs=fs, window=signal.get_window('hamming', 512), noverlap=256, mode='psd')
Ps=np.where(Ps==0, 1e-200 , Ps)
PsdB=10*np.log10(Ps);
[Nf,Nt]=PsdB.shape
#For each audio define the lenght of the unified feature map
multiple=int(np.ceil((Nt/400)))
N=2*multiple-1
lenght=multiple*400
#Create UNIFIED FEATURE MAP
Q=int (lenght/Nt )
rest1 = lenght - Nt*Q
UFeatureMAP[:, 0:(Q*Nt)] = np.tile(PsdB,Q)
UFeatureMAP[:, (Q*Nt):lenght] = PsdB[:, 0:rest1]
#Divide feature map in segments
#Append the segment to list of all segmets and the corrispective to list of labels
for i in range(0,N):
segment=np.copy(UFeatureMAP[: , (i*L):(i*L+M)])
segment=np.expand_dims(segment, axis=2)
list_input.append(segment)
list_label_m.append(label_multi)
list_label_b.append(label_binary)
#Write on a file, for each audio, the number of the segments and the label (binary and multi-class) of the speech
f=open(file_segments,'a',encoding='utf8')
a = '{0}\t{1}\t{2}\n'.format(N,str(label_binary),str(label_multi))
f.write(a)
f.close()
P=len(list_input)
#Each time S audio are processed, the list of the segments is appendend on the file hdf5 and the list is reset
if counter%S==0 and counter!=N_files and P!=0:
print(counter)
dset.resize(dset.shape[0]+P, axis=0)
dset[-P:,:,:,:] = list_input[0:P]
dset_binary.resize(dset_binary.shape[0]+P, axis=0)
dset_binary[-P:] = list_label_b[0:P]
dset_multi.resize(dset_multi.shape[0]+P, axis=0)
dset_multi[-P:] = list_label_m[0:P]
list_input=[]
list_label_m=[]
list_label_b=[]
#This is the last iteration
if counter==N_files and P!=0:
print(counter)
dset.resize(dset.shape[0]+P, axis=0)
dset[-P:,:,:,:] = list_input[0:P]
dset_binary.resize(dset_binary.shape[0]+P, axis=0)
dset_binary[-P:] = list_label_b[0:P]
dset_multi.resize(dset_multi.shape[0]+P, axis=0)
dset_multi[-P:] = list_label_m[0:P]
list_input=[]
list_label_m=[]
list_label_b=[]
fx.close()
fy_multi.close()
fy_binary.close()
print('Done...')
return
def wave_amplitude(data, fs_hz, NFFT, overlap, length, frequenciesRange):
data = np.asarray(data)
data = data[:buffersize, :]
f_eeg_data = filter_data(data, fs_hz)
#t0 = time.time()
if frequenciesRange == 'delta':
wave_band_Hz = np.array([0.4, 4])
elif frequenciesRange == 'theta':
wave_band_Hz = np.array([4, 7])
if frequenciesRange == 'alpha':
wave_band_Hz = np.array([8, 12])
elif frequenciesRange == 'beta':
wave_band_Hz = np.array([12, 25])
elif frequenciesRange == 'gamma':
wave_band_Hz = np.array([25, 35])
size = f_eeg_data.shape[1]
mean_range = np.zeros((1, size))
max_range = np.zeros((1, size))
min_range = np.zeros((1, size))
ratio = np.zeros((1, size))
for channel in range(size):
spec_PSDperHz, freqs, t_spec = mlab.specgram(
f_eeg_data[20:, channel],
NFFT=NFFT,
Fs=fs_hz,
window=mlab.window_hanning,
noverlap=overlap)
# convert the units of the spectral data
spec_PSDperBin = spec_PSDperHz * fs_hz / float(
NFFT) # convert to "Power Spectral Density per bin"
spec_PSDperBin = np.asarray(spec_PSDperBin)
# print(spec_PSDperBin.shape) # from 1 to 110 Hz, step of 1Hz
# take the average spectrum according to the time - axis 1
bool_inds_wave_range = (freqs > wave_band_Hz[0]) & (freqs <
wave_band_Hz[1])
#freq_range = freqs[bool_inds_wave_range == 1]
#freq_range = freq_range[max_range_idx]
spec_PSDperBin_range = spec_PSDperBin[bool_inds_wave_range]
mean_range[0][channel] = np.mean(spec_PSDperBin_range)
# max_range[0][channel] = np.amax(spec_PSDperBin_range)
# get the frequency of the max in each range alpha, beta, theta, gamma
# max_range_idx = np.argmax(spec_PSDperBin_range)
'''
Get the median, max and min of the 4 channels b
'''
med_range = np.median(mean_range[0][:])
max_range = np.amax(mean_range[0][:])
min_range = np.min(mean_range[0][:])
# return [med_alpha, max_alpha, min_alpha, freq_alpha,
# med_beta, max_beta, min_beta, freq_beta,
# med_theta, max_theta, min_theta, freq_theta,
# med_gamma, max_gamma, min_gamma, freq_gamma, time_last_alpha]
results = [med_range, max_range, min_range]
result = results[0]
# print(med_gamma.type())
return result
#print fft_op.shape
#plt.plot(np.arange(44100), np.absolute(np.fft.fft(wav_arr[1][:4096, 0], 44100)))
#plt.show()
print "mean:", np.mean(abs(wav_arr[1][:43000, 0]))
print "median", np.median(abs(wav_arr[1][:43000, 0]))
tmp_arr = [i if i > 350 else 0 for i in wav_arr[1][:, 0]]
write('test.wav', 44100, np.int16(tmp_arr))
filter_op = butter_lowpass_filter(wav_arr[1][:, 0], 1000, 44100, 6)
NFFT = int(fs * 0.1)
noverlap = int(NFFT / 4)
print NFFT, noverlap
print "spec"
Pxx, freqs, t = specgram(wav_arr[1][:43000, 0],
NFFT=NFFT,
window=window_hanning,
noverlap=noverlap)
print Pxx, freqs, t
plt.pcolormesh(t, freqs, Pxx)
plt.colorbar()
plt.show()
Pxx, freqs, t = specgram(filter_op,
NFFT=NFFT,
window=window_hanning,
noverlap=noverlap)
print Pxx, freqs, t
plt.pcolormesh(t, freqs, Pxx)
plt.colorbar()
plt.show()
#print np.amax(butter_lowpass_filter(wav_arr[1][:, 0], 1000, 44100, 6));
from scipy.stats import skew
import numpy as np
#import pandas as pd
# ReadAIFF function
def ReadAIFF(file):
# Reads the frames from the audio clip and returns the uncompressed data
s = aifc.open(file,'r')
nFrames = s.getnframes()
strSig = s.readframes(nFrames)
return np.fromstring(strSig, np.short).byteswap()
# Read one file as an example
params = {'NFFT':256, 'Fs':200, 'noverlap':192}
s = ReadAIFF('train6.aiff')
P, freqs, bins = mlab.specgram(s, **params)
# print(P.shape)
# print(freqs.shape)
# print(bins.shape)
# Spectrogram plotting function long_sample_01.aiff
def plot_spectrogram(ax, P):
plt.imshow(P, origin='lower', extent=[-6,6,-1,1], aspect=4, cmap = cm.get_cmap('bwr'))
loc = plticker.MultipleLocator(base=3.0) # this locator puts ticks at regular intervals
ax.xaxis.set_major_locator(loc)
ax.set_xticklabels(np.arange(1.0,12.5,0.5))
ax.set_yticklabels(range(0,10000,250))
ax.set_xlabel('Time (seconds)', fontsize = 12)
ax.set_ylabel('Frequency (Hz)', fontsize = 12)
cbar = plt.colorbar()
def get_fft_histogram(signal, audio_sample_rate, seconds, histograms_per_second, buckets, plot):
window_size = audio_sample_rate / histograms_per_second
overlap_ratio = 0.0
# FFT the signal and extract frequency components
# Some parameter explanations:
# NFFT = the number of samples grouped into each specgram
# Fs = The sample rate of the signal
specgram = mlab.specgram(
signal,
NFFT=window_size,
Fs=audio_sample_rate,
window=mlab.window_hanning,
noverlap=int(window_size * overlap_ratio))
# The periodogram is a 2D array in the format: [frequencyId][sampleId]
# The actual frequency of each {frequencyId} is recorded in specgram[1]
# where specgram[1][frequencyId] will return the frequency value in Hz
# The {sampleId} represents the sample number after the total number of
# samples is grouped by {window_size}
# Each data point in this array represents signal density at a particular
# frequency and sample number
periodogram = specgram[0]
if verbose:
print "Dimensions of periodogram: %i x %i" % (len(periodogram), len(periodogram[0]))
# apply log transform since specgram() returns linear array
# volume is logrithmic, therefore the periodogram's data (which represents
# amplitude / density of each frequency) must be converted
arr2D = 10 * np.log10(periodogram)
arr2D[arr2D == -np.inf] = 0 # replace infs with zeros
if plot:
fig = plt.figure(figsize=(6, 3.2))
ax = fig.add_subplot(111)
ax.set_title('Spectrogram')
plt.imshow(arr2D)
ax.set_aspect('equal')
plt.colorbar(orientation='vertical')
plt.show()
return None
else:
# Converts periodogram from format [frequencyId][sampleId]
# to [sampleId][frequencyId]
flipped = np.transpose(arr2D)
if verbose:
print "Total samples: %i" % len(flipped)
print "Samples per second: %f" % (len(flipped) / seconds)
print "Grouping FFT into %i-bucket histogram..." % buckets
grouped = []
# Creates histograms for each sample. Groups the entire frequency range
# into {buckets} number of buckets
for i, sample in enumerate(flipped):
perc_done = float(i+1) / len(flipped)
elapsed_seconds = (perc_done * seconds)
histogram = np.array(np.histogram(
sample, bins=buckets)[0]
).tolist()
histogram = [elapsed_seconds] + histogram
grouped.append(histogram)
return grouped
#ax_fs.set_xlim(0.01*max(Pxx),1.1*max(Pxx))
ax_fs.set_xlim(0.01 * max(wavelet.get_wps()), 1.1 * max(wavelet.get_wps()))
pl.setp(ax_ts.get_xticklabels() + ax_fs.get_yticklabels(), visible=False)
pl.tight_layout()
########################
# Fourier Decomposition
from matplotlib import mlab
pl.figure()
S, F, T = mlab.specgram(data,
NFFT=int(0.1 * len(data)),
Fs=sample_rate,
window=mlab.window_hanning,
noverlap=int(.5 * 0.1 * len(data)))
pl.close()
Time = T + min(time)
#collevs=np.linspace(0, max(map(max,S)), 100)
fpeakmin = 2000.0
freq_fourier, Pxx = signal.periodogram(data, fs=sample_rate)
collevs = np.linspace(0, 5 * max(Pxx[freq_fourier > fpeakmin]), 100)
fig, ax_cont = pl.subplots(figsize=(10, 5))
#ax_cont.contourf(Time, F, S, levels=collevs, cmap=cm.gnuplot2)
extent = [min(Time), max(Time), min(F), max(F)]
for i in range(10):
freqs, tempo, Sxx_scipy = signal.spectrogram(sinal, fs=1, nperseg=win * zer_pad, noverlap=(1 - 1 / step_scale) * win)
time1 = time.time()
print('\nSCIPY: {0} ms'.format(100 * (time1 - time0)))
plt.subplot(4, 1, 3)
plt.pcolormesh(tempo, freqs, Sxx_scipy)
plt.ylim(freq_min, freq_max)
plt.plot(tem, w, 'k', lw=2)
plt.ylim(freq_min, freq_max)
plt.title('scipy')
# measure matplotlib spectrogram time
plt.subplot(4, 1, 4)
time0 = time.time()
for i in range(10):
Sxx_mlab, freqs, bins = specgram(sinal, Fs=1, NFFT=win * zer_pad, noverlap=(1 - 1. / step_scale) * win)
time1 = time.time()
print('\nMATPLOTLIB: {0} ms'.format(100 * (time1 - time0)))
print('\ntime resolution:\n CUSTOM: {} \t SCIPY: {} \t MLAB: {}'.format(mat_cs.shape, Sxx_scipy.shape, Sxx_mlab.shape))
print('\ndark line is simulated frequency\n')
print(len(tempo))
plt.plot(tem, w, 'k', lw=2)
plt.pcolormesh(bins, freqs, Sxx_mlab)
plt.ylim(freq_min, freq_max)
plt.title('mlab')
plt.tight_layout()
plt.show()
def data_push(data_url):
"""
Utility function to upload spectogram and wav file from
OOI website to s3.
"""
creds_data = load_creds()
client = boto3.client('s3', aws_access_key_id=creds_data['key_id'],
aws_secret_access_key=creds_data['key_access'])
try:
hydrophone_name = data_url.split('/')[5]
url_date = data_url.split('--')[1][4:14].replace('-','_')
# Read from url
stream = read(data_url)
samp_rate = stream[0].stats.sampling_rate
t_start = stream[0].stats.starttime
t_end = stream[0].stats.endtime
duration = t_end-t_start
# Ping Detections
duration_check = int(duration) - 1
if duration_check > 0:
logging.info('Executing Url: %s', data_url)
pingtimes = np.zeros(duration_check)
for stratpoint in range(duration_check):
pingindex = np.argmax(stream[0].data[int(stratpoint * samp_rate):int((stratpoint + 1) * samp_rate)])
pingtimes[stratpoint] = (t_start + stratpoint + pingindex * stream[0].stats.delta)
# Filter Data+Plot Spectrogram+Save Image and Audio
step_size = 10 # for calculating the rms pressure and ploting the spectrogtam
wlen = 0.056 # bin size in sec
nfft = int(_nearest_pow_2(wlen * samp_rate)) # number of fft points of each bin
per_lap = 0.995 # percentage of overlap
nlap = int(nfft * float(per_lap)) # number of overlapped samples
timestep = 10 # save results every 5 seceonds (no overlap)
for i in range(0, len(pingtimes), timestep):
st = stream.slice(UTCDateTime(pingtimes[i]), UTCDateTime(pingtimes[i]) + step_size)
trace = st[0].copy()
# Plot Spectrogram
npts = len(st[0])
end = npts / samp_rate
# using mlab to create the array of spectrogram
specgram, freq, time = mlab.specgram(trace.data/1e-6, NFFT=nfft, Fs=samp_rate, noverlap=nlap, pad_to=None)
specgram = 10 * np.log10(specgram[1:, :])
specgram = np.flipud(specgram)
freq = freq[1:] / 1e3 # Convert Frequency to kHz
halfbin_time = (time[1] - time[0]) / 2.0
halfbin_freq = (freq[1] - freq[0]) / 2.0
freq = np.concatenate((freq, [freq[-1] + 2 * halfbin_freq]))
time = np.concatenate((time, [time[-1] + 2 * halfbin_time]))
extent = (time[0] - halfbin_time, time[-1] + halfbin_time,
freq[0] - halfbin_freq, freq[-1] + halfbin_freq)
# colormap setting
vmin = 0.50 # default should be 0 to start from the min number of the spectrgram
vmax = 0.95 # default should be 1 to end at the max number of the spectrgram
_range = float(specgram.max() - specgram.min())
vmin = specgram.min() + vmin * _range
vmax = specgram.min() + vmax * _range
norm = Normalize(vmin, vmax) # to scale a 2-D float X input to the (0, 1) range for input to the cmap
# Save spectrogram
fig = plt.figure(frameon=False, figsize=(8, 8))
ax = plt.Axes(fig, [0., 0., 1., 1.])
ax.set_axis_off()
fig.add_axes(ax)
cax = ax.imshow(specgram, interpolation="nearest", extent=extent, norm=norm,
cmap='bone')
dpi = fig.get_dpi()
fig.set_size_inches(512/float(dpi), 512/float(dpi))
ax.axis('tight')
ax.set_xlim(0, end)
ax.set_ylim(0.01, 8)
ax.grid(False)
ax.set_xlabel('Time [s]')
ax.set_ylabel('Frequency [kHz]')
filename = st[0].stats.network+'_'+st[0].stats.station+'_'+st[0].stats.location+'_'+st[0].stats.channel+'_'+str(UTCDateTime(pingtimes[i])).replace("-", "_").replace(
":", "_")
plt.savefig(filename[:-8] + '.jpg')
client.upload_file(filename[:-8] + '.jpg', bucket_name ,'{}/{}/{}/'.format(folder_name, hydrophone_name, url_date) +filename[:-8] + '.jpg')
os.remove(filename[:-8] + '.jpg')
plt.cla()
plt.clf()
plt.close('all')
# save audio
Save2Wav(st[0], filename, samp_rate)
client.upload_file(filename[:-8] + '.wav', bucket_name, '{}/{}/{}/'.format( folder_name, hydrophone_name, url_date) + filename[:-8] + '.wav')
os.remove(filename[:-8] + '.wav')
# delete large objects to release memory.
del trace, st[0]
gc.collect()
else:
logging.info("Skipped Url: %s", data_url)
# check for unwanted URL format.
except Exception as err:
logging.info('Url with error: %s is %s', err, data_url)
def Updated_ReadingAudio(
img_rows=64,
img_cols=64,
path=r"C:\Users\yaniv\Desktop\classification app\goldStandarts_2020_united.xlsx"
):
import scipy
from scipy import signal
from scipy.io import wavfile
from pathlib import Path
import numpy as np
from scipy.misc import imresize
import xlrd
# from AudioPreProcessing import AudioPreProcessing
from skimage import img_as_ubyte
import cv2
import matplotlib.pyplot as plt
from matplotlib import mlab
# Open the gold standard xl
xl_workbook = xlrd.open_workbook(path)
xl_sheet = xl_workbook.sheet_by_index(0)
num_rows = xl_sheet.nrows # Number of rows
Image = [None] * num_rows
TrueLabels = [None] * num_rows
Data = [None] * num_rows
Precentages = [None] * num_rows
for row_idx in range(0, num_rows): # Iterate through rows
current_row = xl_sheet.row_values(row_idx)
data_folder = current_row[0]
file = "%s.wav" % (current_row[1]) #audiofile
file_to_open = data_folder + file #audiofile
Fs, samples = wavfile.read(file_to_open)
times1 = int(
round(current_row[2], 4) *
Fs) #taking relevant times of the syllable from the audio file
times2 = int(
round(current_row[3], 4) *
Fs) #taking relevant times of the syllable from the audio file
Signal = samples[
times1:
times2] #taking relevant times of the syllable from the audio file
spectrogram, frequencies, times = mlab.specgram(
Signal, NFFT=256, Fs=Fs,
noverlap=120) # same spectrogram as the matlab
# dBS = 20*np.log10(spectrogram) # convert to dB
# plt.figure()
# plt.pcolormesh(dBS)
# spectrogram(sig,256,120,256,Fs,'MinThreshold',-110,'yaxis')
# data = PreProcessedAudio #Dummy data. Just for testing
width = img_rows # define the new spectrogram shape
height = img_cols
dim = (width, height)
resized = cv2.resize(spectrogram, dim, interpolation=cv2.INTER_AREA)
# plt.figure()
# plt.pcolormesh(20*np.log10(resized))
# plt.colorbar()
# img = resized/np.amax(resized)
# plt.figure()
# plt.pcolormesh(20*np.log10(img))
# plt.colorbar()
Data[row_idx] = resized
TrueLabels[row_idx] = current_row[8]
return (Data, TrueLabels)
def rawPeakSPD(x, t=None, NFFT=1024): # A power, not amplitude
return(np.max(specgram(x, NFFT=NFFT, noverlap=0)[0].flatten()))
def digital_to_spec(
digital: np.ndarray,
fs: float,
frac_cut: float,
plot: bool = False
) -> Union[Tuple[np.ndarray, float], Tuple[np.ndarray, float, Figure, Axes,
float, float]]:
"""Produces a spectrogram and a cut-off intensity to yield the
specified fraction of data.
Parameters
----------
digital : numpy.ndarray, shape=(Ts, )
The sampled audio-signal.
fs : float
The sample-frequency used to create the digital signal.
frac_cut : float
The fractional portion of intensities for which the cutoff is selected.
E.g. frac_cut=0.8 will produce a cutoff intensity such that the bottom 80%
of intensities are excluded.
plot : bool
If True, produce a plot of the spectrogram and return the
matplotlib fig & ax objects.
Returns
-------
Union[Tuple[numpy.ndarray, float]]
The spectrogram and the desired cutoff
If plot=True, then: (spectrogram, cutoff, fig, ax, df, dt)
is returned. Where (fig, ax) are the plot objects, and df
and dt are the frequency and time units associated with the
spectrogram bins."""
if digital.max() <= 1:
digital = digital * 2**15
assert 0.0 <= frac_cut <= 1.0
kwargs = dict(NFFT=4096,
Fs=fs,
window=mlab.window_hanning,
noverlap=int(4096 / 2))
if not plot:
S, freqs, times = mlab.specgram(digital, **kwargs)
else:
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
S, freqs, times, im = ax.specgram(digital, **kwargs)
fig.colorbar(im)
# log-scaled Fourier amplitudes have a much more gradual distribution
# for audio data.
np.clip(S, a_min=1e-20, a_max=None, out=S)
np.log(S, out=S)
# Compute the cumulative distribution over Fourier component log-amplitudes.
# Use this to identify the threshold amplitude below which `frac_cutoff`
# proportion of amplitudes lie.
a = np.sort(S.flatten())
cutoff = a[int(len(a) * frac_cut)]
if not plot:
return S, cutoff
else:
df = freqs[1] - freqs[0]
dt = times[1] - times[0]
return S, cutoff, fig, ax, df, dt
def recomputeSpectrogram(self, indexFrom=None, indexTo=None, maxCol=None):
"""
Method that computes the spectrogram of the signal in the supplied interval
:param indexFrom: Start index in signal data array coordinates
:param indexTo: End index in signal data array coordinates
"""
# do not work fine if the signal data changes
# if not self.changes and self.lastInterval == (indexFrom, indexTo):
# return
indexFrom = indexFrom if indexFrom is not None else self.lastInterval[0]
indexTo = indexTo if indexTo is not None else self.lastInterval[1]
indexTo = indexTo if indexTo != -1 else self.signal.length
maxCol = maxCol if maxCol != 0 else self.signal.length / self.NFFT
if self.signal is None:
raise Exception("No signal to compute spectrogram")
# computing overlap so the number of columns is less or equal than maxCol
smin = indexFrom - self.NFFT
smax = indexTo + self.NFFT
pre, post = np.zeros(max(-smin, 0)), np.zeros(
max(smax - self.signal.length, 0))
data = np.concatenate(
(pre, self.signal.data[max(smin, 0):min(smax, self.signal.length)],
post))
self.__visual_overlap = self.overlap
if maxCol is not None:
cs = self.NFFT - self.overlap
cols = (smax - smin - self.overlap) / cs
if cols > maxCol:
self.__visual_overlap = int(self.NFFT - (smax - smin) / maxCol)
if False and indexFrom == self.lastInterval[
0] and indexTo == self.lastInterval[1]:
return
# delegate in matplotlib specgram function the process of compute the spectrogram
self.matriz, self.freqs, self.bins = mlab.specgram(
data,
NFFT=self.NFFT,
Fs=self.signal.samplingRate,
detrend=mlab.detrend_none,
window=self.window,
noverlap=self.__visual_overlap,
sides=self.SPECGRAM_COMPLEX_SIDE)
# changes to dB for bio acoustic processing
temp = np.amax(self.matriz)
if temp == 0:
temp = 1.0
self.matriz = 10. * np.log10(self.matriz / temp)
Zfin = np.isfinite(self.matriz)
if np.any(Zfin):
m = self.matriz[Zfin].min()
self.matriz[np.isneginf(self.matriz)] = m
else:
self.matriz[self.matriz <
-100] = -100 # verificar estos valores constantes
self.matriz = np.transpose(self.matriz)
self.lastInterval = (indexFrom, indexTo)
self.lastMaxCol = maxCol
self.changes = False
rcv_file.write(RecivedData)
print("reciving another bounch of data")
myconn.settimeout(5.0)
RecivedData = myconn.recv(BUFF_SIZE)
print(RecivedData)
rcv_file.close()
print("\n File has been copied successfully \n")
myconn.close()
print("\n Server closed the connection \n")
server_socket.close()
print("\n Plotting... \n")
acc_data = np.genfromtxt(rcv_filename, delimiter=',', names=True)
acc_x, freq_x, _ = mlab.specgram(acc_data['x'],
Fs=SAMPLE_RATE,
NFFT=SAMPLE_RATE * SMAPLE_TIME)
acc_y, freq_y, _ = mlab.specgram(acc_data['y'],
Fs=SAMPLE_RATE,
NFFT=SAMPLE_RATE * SMAPLE_TIME)
acc_z, freq_z, _ = mlab.specgram(acc_data['z'],
Fs=SAMPLE_RATE,
NFFT=SAMPLE_RATE * SMAPLE_TIME)
plt.plot(freq_x[10:], acc_x[10:], label='x', linewidth=0.5)
plt.plot(freq_y[10:], acc_y[10:], label='y', linewidth=0.5)
plt.plot(freq_z[10:], acc_z[10:], label='z', linewidth=0.5)
plt.yscale('log')
plt.xlim((0, 160))
plt.legend(loc='upper right')
plt.show()
plt.savefig('spectrum.png')
def update_plots(self, draw=True):
self.remove_peak_annotation()
# trace:
self.axt.set_xlim(self.toffset, self.toffset + self.twindow)
t0 = int(np.round(self.toffset * self.samplerate))
t1 = int(np.round((self.toffset + self.twindow) * self.samplerate))
if t1 > len(self.data):
t1 = len(self.data)
time = np.arange(t0, t1) / self.samplerate
if self.trace_artist == None:
self.trace_artist, = self.axt.plot(time, self.data[t0:t1])
else:
self.trace_artist.set_data(time, self.data[t0:t1])
self.axt.set_ylim(self.ymin, self.ymax)
# compute power spectrum:
nfft = int(
np.round(2**(np.floor(
np.log(self.samplerate / self.fresolution) / np.log(2.0)) +
1.0)))
if nfft < 16:
nfft = 16
nfft, noverlap = nfft_noverlap(self.fresolution, self.samplerate, 0.5,
16)
t00 = t0
t11 = t1
w = t11 - t00
minw = nfft * (self.cfg['minPSDAverages'][0] + 1) // 2
if t11 - t00 < minw:
w = minw
t11 = t00 + w
if t11 >= len(self.data):
t11 = len(self.data)
t00 = t11 - w
if t00 < 0:
t00 = 0
t11 = w
power, freqs = ml.psd(self.data[t00:t11],
NFFT=nfft,
noverlap=noverlap,
Fs=self.samplerate,
detrend=ml.detrend_mean)
self.deltaf = freqs[1] - freqs[0]
# detect fish:
h_kwargs = psd_peak_detection_args(self.cfg)
h_kwargs.update(harmonic_groups_args(self.cfg))
self.fishlist, fzero_harmonics, self.mains, self.allpeaks, peaks, lowth, highth, center = harmonic_groups(
freqs, power, verbose=self.verbose, **h_kwargs)
highth = center + highth - 0.5 * lowth
lowth = center + 0.5 * lowth
# spectrogram:
t2 = t1 + nfft
specpower, freqs, bins = ml.specgram(self.data[t0:t2],
NFFT=nfft,
Fs=self.samplerate,
noverlap=nfft // 2,
detrend=ml.detrend_mean)
z = decibel(specpower)
z = np.flipud(z)
extent = self.toffset, self.toffset + np.amax(
bins), freqs[0], freqs[-1]
self.axs.set_xlim(self.toffset, self.toffset + self.twindow)
if self.spectrogram_artist == None:
self.fmax = np.round((freqs[-1] / 4.0) / 100.0) * 100.0
min = highth
min = np.percentile(z, 70.0)
max = np.percentile(z, 99.9) + 30.0
# cm = plt.get_cmap( 'hot_r' )
cm = plt.get_cmap('jet')
self.spectrogram_artist = self.axs.imshow(z,
aspect='auto',
extent=extent,
vmin=min,
vmax=max,
cmap=cm,
zorder=1)
else:
self.spectrogram_artist.set_data(z)
self.spectrogram_artist.set_extent(extent)
self.axs.set_ylim(self.fmin, self.fmax)
# power spectrum:
self.axp.set_xlim(self.fmin, self.fmax)
if self.deltaf >= 1000.0:
dfs = '%.3gkHz' % 0.001 * self.deltaf
else:
dfs = '%.3gHz' % self.deltaf
tw = float(w) / self.samplerate
if tw < 1.0:
tws = '%.3gms' % (1000.0 * tw)
else:
tws = '%.3gs' % (tw)
a = 2 * w // nfft - 1 # number of ffts
m = ''
if self.cfg['mainsFreq'][0] > 0.0:
m = ', mains=%.0fHz' % self.cfg['mainsFreq'][0]
if self.power_frequency_label == None:
self.power_frequency_label = self.axp.set_xlabel(
r'Frequency [Hz] (nfft={:d}, $\Delta f$={:s}: T={:s}/{:d}{:s})'
.format(nfft, dfs, tws, a, m))
else:
self.power_frequency_label.set_text(
r'Frequency [Hz] (nfft={:d}, $\Delta f$={:s}: T={:s}/{:d}{:s})'
.format(nfft, dfs, tws, a, m))
self.axp.set_xlim(self.fmin, self.fmax)
if self.power_label == None:
self.power_label = self.axp.set_ylabel('Power')
if self.decibel:
if len(self.allpeaks) > 0:
self.allpeaks[:, 1] = decibel(self.allpeaks[:, 1])
power = decibel(power)
pmin = np.min(power[freqs < self.fmax])
pmin = np.floor(pmin / 10.0) * 10.0
pmax = np.max(power[freqs < self.fmax])
pmax = np.ceil(pmax / 10.0) * 10.0
doty = pmax - 5.0
self.power_label.set_text('Power [dB]')
self.axp.set_ylim(pmin, pmax)
else:
pmax = np.max(power[freqs < self.fmax])
doty = pmax
pmax *= 1.1
self.power_label.set_text('Power')
self.axp.set_ylim(0.0, pmax)
if self.all_peaks_artis == None:
self.all_peaks_artis, = self.axp.plot(
self.allpeaks[:, 0],
np.zeros(len(self.allpeaks[:, 0])) + doty,
'o',
color='#ffffff')
self.good_peaks_artist, = self.axp.plot(peaks,
np.zeros(len(peaks)) +
doty,
'o',
color='#888888')
else:
self.all_peaks_artis.set_data(
self.allpeaks[:, 0],
np.zeros(len(self.allpeaks[:, 0])) + doty)
self.good_peaks_artist.set_data(peaks, np.zeros(len(peaks)) + doty)
labels = []
fsizes = [
np.sqrt(np.sum(self.fishlist[k][:, 1]))
for k in range(len(self.fishlist))
]
fmaxsize = np.max(fsizes) if len(fsizes) > 0 else 1.0
for k in range(len(self.peak_artists)):
self.peak_artists[k].remove()
self.peak_artists = []
for k in range(len(self.fishlist)):
if k >= len(self.markerrange):
break
fpeaks = self.fishlist[k][:, 0]
fpeakinx = [
int(np.round(fp / self.deltaf)) for fp in fpeaks
if fp < freqs[-1]
]
fsize = 7.0 + 10.0 * (fsizes[k] / fmaxsize)**0.5
fishpoints, = self.axp.plot(
fpeaks[:len(fpeakinx)],
power[fpeakinx],
linestyle='None',
color=self.colorrange[k % len(self.colorrange)],
marker=self.markerrange[k],
ms=fsize,
mec=None,
mew=0.0,
zorder=1)
self.peak_artists.append(fishpoints)
if self.deltaf < 0.1:
labels.append('%4.2f Hz' % fpeaks[0])
elif self.deltaf < 1.0:
labels.append('%4.1f Hz' % fpeaks[0])
else:
labels.append('%4.0f Hz' % fpeaks[0])
if len(self.mains) > 0:
fpeaks = self.mains[:, 0]
fpeakinx = [
np.round(fp / self.deltaf) for fp in fpeaks if fp < freqs[-1]
]
fishpoints, = self.axp.plot(fpeaks[:len(fpeakinx)],
power[fpeakinx],
linestyle='None',
marker='.',
color='k',
ms=10,
mec=None,
mew=0.0,
zorder=2)
self.peak_artists.append(fishpoints)
labels.append('%3.0f Hz mains' % self.cfg['mainsFreq'][0])
ncol = (len(labels) - 1) // 8 + 1
self.legendhandle = self.axs.legend(self.peak_artists[:len(labels)],
labels,
loc='upper right',
ncol=ncol)
self.legenddict = dict()
for legpoints, (finx, fish) in zip(self.legendhandle.get_lines(),
enumerate(self.fishlist)):
legpoints.set_picker(8)
self.legenddict[legpoints] = [finx, fish]
self.legendhandle.set_visible(self.legend)
if self.power_artist == None:
self.power_artist, = self.axp.plot(freqs, power, 'b', zorder=3)
else:
self.power_artist.set_data(freqs, power)
if draw:
self.fig.canvas.draw()
def make(self,
raw_audio,
samp_freq):
"""makes spectrogram using assigned properties
Parameters
----------
raw_audio : 1-d numpy array
raw audio waveform
samp_freq : integer scalar
sampling frequency in Hz
Returns
-------
spect : 2-d numpy array
freq_bins : 1-d numpy array
time_bins : 1-d numpy array
"""
if self.filterFunc == 'diff':
raw_audio = np.diff(raw_audio) # differential filter_func, as applied in Tachibana Okanoya 2014
elif self.filterFunc == 'bandpass_filtfilt':
raw_audio = hvc.evfuncs.bandpass_filtfilt(raw_audio,
samp_freq,
self.freqCutoffs)
elif self.filterFunc == 'butter_bandpass':
raw_audio = butter_bandpass_filter(raw_audio,
samp_freq,
self.freqCutoffs)
try: # try to make spectrogram
if self.spectFunc == 'scipy':
if self.window is not None:
freq_bins, time_bins, spect = scipy.signal.spectrogram(raw_audio,
samp_freq,
window=self.window,
nperseg=self.nperseg,
noverlap=self.noverlap)
else:
freq_bins, time_bins, spect = scipy.signal.spectrogram(raw_audio,
samp_freq,
nperseg=self.nperseg,
noverlap=self.noverlap)
elif self.spectFunc == 'mpl':
# note that the matlab specgram function returns the STFT by default
# whereas the default for the matplotlib.mlab version of specgram
# returns the PSD. So to get the behavior of matplotlib.mlab.specgram
# to match, mode must be set to 'complex'
# I think I determined empirically at one point (by staring at single
# cases) that mlab.specgram gave me values that were closer to Matlab's
# specgram function than scipy.signal.spectrogram
# Matlab's specgram is what Tachibana used in his original feature
# extraction code. So I'm maintaining the option to use it here.
# 'mpl' is set to return complex frequency spectrum,
# not power spectral density,
# because some tachibana features (based on CUIDADO feature set)
# need to use the freq. spectrum before taking np.abs or np.log10
if self.window is not None:
spect, freq_bins, time_bins = specgram(raw_audio,
NFFT=self.nperseg,
Fs=samp_freq,
window=self.window,
noverlap=self.noverlap,
mode='complex')
else:
spect, freq_bins, time_bins = specgram(raw_audio,
NFFT=self.nperseg,
Fs=samp_freq,
noverlap=self.noverlap,
mode='complex')
except ValueError as err: # if `try` to make spectrogram raised error
if str(err) == 'window is longer than input signal':
raise WindowError()
else: # unrecognized error
raise
if self.remove_dc:
# remove zero-frequency component
freq_bins = freq_bins[1:]
spect = spect[1:,:]
# we take the absolute magnitude
# because we almost always want just that for our purposes
spect = np.abs(spect)
if self.logTransformSpect:
spect = np.log10(spect) # log transform to increase range
if self.thresh is not None:
spect[spect < self.thresh] = self.thresh
# below, I set freq_bins to >= freq_cutoffs
# so that Koumura default of [1000,8000] returns 112 freq. bins
if self.freqCutoffs is not None:
f_inds = np.nonzero((freq_bins >= self.freqCutoffs[0]) &
(freq_bins <= self.freqCutoffs[1]))[0] # returns tuple
freq_bins = freq_bins[f_inds]
spect = spect[f_inds, :]
return spect, freq_bins, time_bins
def make_spectrogram(self, raw_audio, chunk_size=1000):
'''
Given data, compute a spectrogram. To avoid slow memory
issues build the spectrogram in a stream-like fashion
where we build it in chunk sizes of 1000 spectrogram frames
Assumptions:
o self.framerate contains the data framerate
o self.nfft contains the window size
o self.hop contains the hop size for fft
o self.pad_to contains the zero-padding that
we add if self.pad_to > self.nfft
Returns a two-tuple: an array of segment times
and a 2D spectrogram array
@param data: the time/amplitude data
@type data: np.array([float])
@param chunk_size: controls the incremental size used to
build up the spectrogram
@type chunk_size: int
@return: (time slices arrya, spectrogram)
@rtype: (np.array([float]), np.array([float]))
'''
# The time_labels will be in seconds. But
# they will be fractions of a second, like
# 0.256, ... 1440
self.log.info("Creating spectrogram...")
# Compute the number of raw audio frames
# needed to generate chunk_size spec frames
len_chunk = (chunk_size - 1) * self.hop + self.nfft
final_spec = None
slice_times = None
start_chunk = 0
# Generate 1000 spect frames at a time, being careful to follow the correct indexing at the boarders to
# "simulate" the full fft. Namely, if we use indeces (raw_start, raw_end) to get the raw_audio frames needed
# to generate the spectrogram chunk, remember the next chunk does not start at raw_end but actually start
# and (raw_end - NFFT) + hop. **THIS IS KEY** to propertly simulate the full spectrogram creation process
iteration = 0
print("Approx number of chunks:", int(raw_audio.shape[0] / len_chunk))
while start_chunk + len_chunk < raw_audio.shape[0]:
if (iteration % 100 == 0):
print("Chunk number " + str(iteration))
[spectrum, freqs,
t] = ml.specgram(raw_audio[start_chunk:start_chunk + len_chunk],
NFFT=self.nfft,
Fs=self.framerate,
noverlap=(self.nfft - self.hop),
window=ml.window_hanning,
pad_to=self.pad_to)
# Cutout uneeded high frequencies!
spectrum = spectrum[(freqs <= self.max_freq)]
if start_chunk == 0:
final_spec = spectrum
slice_times = t
else:
final_spec = np.concatenate((final_spec, spectrum), axis=1)
# Shift t to be 0 started than Offset the new times
# by the last frame's time + the time gap between frames (= hop / fr)
t = t - t[0] + slice_times[-1] + (self.hop / self.framerate)
slice_times = np.concatenate((slice_times, t))
# Remember that we want to start as if we are doing one continuous sliding window
start_chunk += len_chunk - self.nfft + self.hop
iteration += 1
# Do one final chunk for whatever remains at the end
[spectrum, freqs, t] = ml.specgram(raw_audio[start_chunk:],
NFFT=self.nfft,
Fs=self.framerate,
noverlap=(self.nfft - self.hop),
window=ml.window_hanning,
pad_to=self.pad_to)
# Cutout the high frequencies that are not of interest
spectrum = spectrum[(freqs <= self.max_freq)]
final_spec = np.concatenate((final_spec, spectrum), axis=1)
# Update the times:
t = t - t[0] + slice_times[-1] + (self.hop / self.framerate)
slice_times = np.concatenate((slice_times, t))
# check the shape of this
self.log.info("Done creating spectrogram.")
# This we may actually want to do! Log transform!!!
# Transformer for magnitude to power dB:
#amp_to_dB_transformer = torchaudio.transforms.AmplitudeToDB()
#freq_time_dB_tensor = amp_to_dB_transformer(torch.Tensor(freq_time))
# Note transpose the spectrogram to be of shape - (time, freq)
return final_spec.T, slice_times
LEEG3_selection[0].T) #plots on top of previous plot with same y axis
# Plotting data with channel names
channel_names = ['L_EEG_3', 'R_EEG_4']
two_eeg_chans = raw[channel_names, start_sample:stop_sample]
y_offset = np.array([1e-4, 0]) # just enough to separate the channel traces
x = two_eeg_chans[1]
y = two_eeg_chans[0].T + y_offset
lines = plt.plot(x, y)
plt.legend(lines, channel_names)
#%%
# Plotting spectrogram for first 50 seconds of first channel
NFFT = 1024 # the length of the windowing segments
spectrum1, freqs1, bins = mlab.specgram(LEEG3_selection[0][0],
NFFT=NFFT,
Fs=Fs,
noverlap=900)
spectrum2, freqs2, bins = mlab.specgram(REEG4_selection[0][0],
NFFT=NFFT,
Fs=Fs,
noverlap=900)
# Only run for [0,5000]
min_val = 5 * np.log10(min(spectrum1.min(), spectrum2.min()))
max_val = 15.5 * np.log10(min(spectrum1.max(), spectrum2.max()))
start_stop_seconds = np.array([0, 5000])
start_sample, stop_sample = (start_stop_seconds * Fs).astype(int)
cmap = plt.get_cmap('magma')
cmap.set_under(color='k', alpha=None)
ax1 = plt.subplot(311)
t_sec = np.array(range(0, f_eeg_data_uV.size)) / fs_Hz
plt.plot(t_sec, f_eeg_data_uV)
plt.ylim(-100, 100)
plt.ylabel('EEG (uV)')
plt.xlabel('Time (sec)')
plt.title(fname)
plt.xlim(t_sec[0], t_sec[-1])
# make spectrogram
ax = plt.subplot(312, sharex=ax1)
overlap = NFFT - int(0.25 * fs_Hz) # quarter-second steps
spec_PSDperHz, freqs, t_spec = mlab.specgram(
np.squeeze(f_eeg_data_uV),
NFFT=NFFT,
window=mlab.window_hanning,
Fs=fs_Hz,
noverlap=overlap) # returns PSD power per Hz
spec_PSDperBin = spec_PSDperHz * fs_Hz / float(NFFT) #convert to "per bin"
del spec_PSDperHz # remove this variable so that I don't mistakenly use it
#reduce size
full_spec_PSDperBin = spec_PSDperBin
full_t_spec = t_spec
spec_PSDperBin = full_spec_PSDperBin[:, 1:-1:2] # get every other time slice
t_spec = full_t_spec[1:-1:2] # get every other time slice
plt.pcolor(t_spec, freqs, 10 * np.log10(spec_PSDperBin)) # dB re: 1 uV
plt.clim(25 - 5 + np.array([-40, 0]))
plt.xlim(t_sec[0], t_sec[-1])
if (t_lim_sec[2 - 1] != 0):
def plot_spectrogram(st,
start_sec,
end_sec,
vmin,
vmax,
fkhz_min,
fkhz_max,
adcp=''):
def _nearest_pow_2(x):
a = M.pow(2, M.ceil(np.log2(x)))
b = M.pow(2, M.floor(np.log2(x)))
if abs(a - x) < abs(b - x):
return a
else:
return b
spec_data = st.slice(st[0].stats.starttime + start_sec,
st[0].stats.starttime + end_sec)
fs = st[0].stats.sampling_rate
wlen = 0.056
# bin size in sec
npts = len(spec_data[0])
end = npts / fs
nfft = int(_nearest_pow_2(wlen * fs)) # number of fft points of each bin
print(nfft)
per_lap = 0.10 # percentage of overlap
nlap = int(nfft * float(per_lap)) # number of overlapped samples
# remove ADCP
if adcp == 'energy':
samples = remove_adcp_energy(spec_data[0].data, nfft)
#elif adcp == 'period':
# samples = remove_adcp_periodicity(spec_data[0].data, fs)
else:
samples = spec_data[0].data
# using mlab to create the array of spectrogram
specgram, freq, time = mlab.specgram(samples,
NFFT=nfft,
Fs=fs,
noverlap=nlap,
pad_to=None)
specgram = 10 * np.log10(specgram[1:, :]) + 169 - 128.9
specgram = np.flipud(specgram)
freq = freq[1:] / 1e3 # Convert Frequency to kHz
halfbin_time = (time[1] - time[0]) / 2.0
halfbin_freq = (freq[1] - freq[0]) / 2.0
freq = np.concatenate((freq, [freq[-1] + 2 * halfbin_freq]))
time = np.concatenate((time, [time[-1] + 2 * halfbin_time]))
extent = (time[0] - halfbin_time, time[-1] + halfbin_time,
freq[0] - halfbin_freq, freq[-1] + halfbin_freq)
# colormap setting
#vmin = 0.50 # default should be 0 to start from the min number of the spectrgram
#vmax = 0.90 # default should be 1 to end at the max number of the spectrgram
_range = float(specgram.max() - specgram.min())
vmin = specgram.min() + vmin * _range
vmax = specgram.min() + vmax * _range
norm = Normalize(
vmin, vmax
) # to scale a 2-D float X input to the (0, 1) range for input to the cmap
# plot spectrogram
fig = plt.figure(figsize=(8, 8))
ax = fig.add_subplot(111)
cax = ax.imshow(specgram,
interpolation="nearest",
extent=extent,
norm=norm,
cmap='viridis')
dpi = fig.get_dpi()
fig.set_size_inches(512 / float(dpi), 512 / float(dpi))
ax.axis('tight')
ax.set_xlim(0, end)
ax.set_ylim(fkhz_min, fkhz_max)
ax.grid(False)
ax.set_xlabel('Time [s]')
ax.set_ylabel('Frequency [kHz]')
ax.set_title(spec_data[0].stats.starttime)
cbar = fig.colorbar(cax)
dict_995 = np.load('../Ads_fingerprint_99.5.npy').item()
while 1 != 2:
sleep(2)
#onlyfiles = [f for f in os.listdir() if os.path.isfile(str(f))]
time_stamp_list = []
for k in os.listdir(): #onlyfiles:
time_stamp_list.append(time.ctime(os.path.getmtime(str(k))))
Latest_Stream = os.listdir()[time_stamp_list.index(max(time_stamp_list))]
rate1, song_array1 = wavfile.read(Latest_Stream)
#rate2, song_array2 = wavfile.read('Josh-Woodward--I-Want-To-Destroy-Something-Beautiful_21sec_cut.wav')
spec1, freqs1, t1 = specgram(song_array1[:, 0],
NFFT=4096,
Fs=rate1,
noverlap=2048)
#spec2, freqs2, t2 = specgram(song_array2[:,0], NFFT=4096, Fs=rate2, noverlap=2048)
spec1[spec1 == 0] = 1e-6
#spec2[spec2 == 0] = 1e-6
min_freq = 0
max_freq = 15000
Z1, freqs1 = cut_specgram(min_freq, max_freq, spec1, freqs1)
#Z2, freqs2 = cut_specgram(min_freq, max_freq, spec2, freqs2)
coordinates1 = peak_local_max(Z1, min_distance=20, threshold_abs=20)
#coordinates2 = peak_local_max(Z2, min_distance=20, threshold_abs=20)
#len(song_array2)/len(song_array1)
ad_vec = 0
tvec = {}
for ix in dict_995.keys():
idx = 0
tock()
# normalise
if settings['normalise_volume']:
print "Normalise volume"
X_downsampled -= X_downsampled.mean(1).reshape(-1, 1)
X_downsampled /= X_downsampled.std(1).reshape(-1, 1)
tock()
# compute spectrograms
print "Compute spectrograms"
nfft = settings['specgram_num_components']
noverlap = nfft * (1 - 1. / settings['specgram_redundancy'])
log_scale = settings['log_scale']
dummy = specgram(X_downsampled[0], NFFT=nfft,
noverlap=noverlap)[0] # to get the dimensions
X_specgram = np.zeros((X.shape[0], dummy.shape[0], dummy.shape[1]),
dtype=X.dtype)
for k in xrange(X.shape[0]):
X_specgram[k] = specgram(X_downsampled[k], NFFT=nfft, noverlap=noverlap)[0]
X_specgram = np.log(1 + log_scale * X_specgram)
tock()
if 'lnorm' in settings and settings[
'lnorm']: # check presence first for backwards compatibility
print "Locally normalise spectrograms"
lnorm_sigma_mean = settings['lnorm_sigma_mean']
lnorm_sigma_std = settings['lnorm_sigma_std']
def get_specgram(data):
arr2D, freqs, bins = specgram(data, window=window_hanning,
Fs=fs_Hz, NFFT=nfft, noverlap=overlap)
return arr2D, freqs, bins
print 'Data: {} samples'.format(data.size)
print 'Data Min/Max: ',data.min(),data.max()
data = data << (16 - NBPS)
data = data >> (16 - NBPS)
print 'Data Min/Max: ',data.min(),data.max()
if data.min() <= -2**(NBPS-1) : print 'Warning: MIN Saturated !!!!!'
elif data.min() < -2**(NBPS-1)+16 : print 'Note: MIN close to be saturated !!!!!'
if data.max() >= 2**(NBPS-1)-1 : print 'Warning: MAX Saturated !!!!!'
elif data.max() > 2**(NBPS-1)-16 : print 'Note: MAX close to be saturated !!!!!'
data = data.astype(float) * (2**-(NBPS-1)); #scale to 0dBFS
data = np.multiply(data, np.cos((np.arange(data.size)*2*np.pi*120.0/240.0)+0.06287))
spectrum, unused1, unused2 = mlab.specgram(data, NFFT=NFFT, window = np.blackman(NFFT), noverlap=NFFT*overlap)
spectrum /= NFFT/8;
print 'Spectrogram shape: {0[0]}:{0[1]}'.format(spectrum.shape)
cum_hist = spec_hist(spectrum, NFFT=NFFT, Pmin=Pmin, aspect=aspect)
im = Image.fromarray( norm255(np.log2(cum_hist)) )
im.putpalette(getpalette(plt.cm.gnuplot2))
im = im.convert('RGBA').rotate(90)
bg = Image.new('RGBA', im.size, (0, 0, 0, 0))
draw = ImageDraw.Draw(bg)
FONTSIZE = 12
font = ImageFont.truetype("arial.ttf", FONTSIZE)
for i in xrange(Pmax, FS_span, 5): # --- power level grid
y = i*(float(bg.size[1])/FS_span)
def a2(x,
fm,
plim=None,
dt=None,
n=None,
tw=None,
dlog2p=None,
s0=None): # subharmonic summation
#VARIABLES
if plim is None: # rango de los candidatos a altura
plim = [50, 800]
if dt is None: # tiempo entre estimados de altura
dt = 0.010
if n is None: # cantidad de armonicas a analizar
n = 7
if tw is None: # tamano de la ventana (segundos)
tw = 0.050
if dlog2p is None: # distancia entre los candidatos (octavas)
dlog2p = 1 / 48
if s0 is None: # umbral de claridad de altura
s0 = 0
# CANDIDATES
log2pc = arange(log2(plim[0]), log2(plim[1]),
dlog2p) # log2 de los candidatos a altura
pc = 2**log2pc # candidatos a altura
# VENTANA
ws = 2**round(log2(tw * fm)) # tamano de la ventana (muestras)
w = hann(ws)
# ventana Hann de tamano ws
o = ws / 2
# traslape entre las ventanas (recomendado para ventana Hann: 50%)
# SIGNAL
t = arange(0, len(x) / fm, dt) # tiempos de los estimados de altura
x = concatenate(
(zeros(ws // 2), x, zeros(ws // 2))
) # se agregan ceros al principio para garantizar que ventanas cubran principio y final de la senal
S = zeros(
(len(pc),
len(t))) # matriz de puntos de los candidatos a lo largo del tiempo
(X, f, tj) = specgram(
x, ws, fm, window=w, noverlap=o, mode='magnitude'
) # calculo del espectro; devuelve tambien frecuencias y tiempos
# INTERATE OVER THE CANDIDATES
for i in range(0, len(pc)): # para cada candidato:
fi = pc[i] * arange(1, n + 1) # identifique primeras n armonicas
A = zeros((len(fi), X.shape[1])) # cree matriz nArmonicas x nVentanas
for j in range(0, X.shape[1]): # por cada ventana:
A[:, j] = interp(fi, f,
X[:, j]) # estime el espectro en las armonicas
si = sum(
A, 0
) # calcule el puntaje del candidato en cada ventana multiplicando el valor del espectro en las armonicas
S[i, :] = interp(t, tj,
si) # interpole los puntajes a los tiempos deseados
p = pc[argmax(
S, 0)] # escoja el candidato con mayor puntuacion en cada ventana
s = amax(S, 0) # registre la puntuacion obtenida
p[s < s0] = float(
'nan') # # indefina alturas cuyo puntaje no excede el umbral
return (
p, t, s, pc, S
) # devuelva ganadores, tiempos, puntajes, lista de candidatos y matriz de puntajes
def spectrogram1(data,
samp_rate,
per_lap=0.9,
wlen=None,
log=False,
fmt=None,
dbscale=False,
y_lim=False,
title=False,
fig_mark=[0],
xtime_offset=0,
mult=8.0,
axes=False,
zorder=None,
clip=[0.0, 1.0]):
samp_rate = float(samp_rate)
if not wlen:
wlen = samp_rate / 100.0
npts = len(data)
nfft = int(_nearest_pow_2(wlen * samp_rate))
if nfft > npts:
nfft = int(_nearest_pow_2(npts / 8.0))
if mult is not None:
mult = int(_nearest_pow_2(mult))
mult = mult * nfft
nlap = int(nfft * float(per_lap))
# data = data - data.mean()
end = npts / samp_rate
specgram, freq, time = mlab.specgram(data,
Fs=samp_rate,
NFFT=nfft,
pad_to=mult,
noverlap=nlap)
if dbscale:
specgram = 10 * np.log10(specgram[1:, :])
else:
specgram = np.sqrt(specgram[1:, :])
freq = freq[1:]
vmin, vmax = clip
if vmin < 0 or vmax > 1 or vmin >= vmax:
msg = "Invalid parameters for clip option."
raise ValueError(msg)
_range = float(specgram.max() - specgram.min())
vmin = specgram.min() + vmin * _range
vmax = specgram.min() + vmax * _range
norm = Normalize(vmin, vmax, clip=True)
if not axes:
fig = plt.figure()
ax = fig.add_subplot(111)
else:
ax = axes
halfbin_time = (time[1] - time[0]) / 2.0
halfbin_freq = (freq[1] - freq[0]) / 2.0
if log:
# pcolor expects one bin more at the right end
freq = np.concatenate((freq, [freq[-1] + 2 * halfbin_freq]))
time = np.concatenate((time, [time[-1] + 2 * halfbin_time]))
# center bin
time -= halfbin_time
freq -= halfbin_freq
time = time + xtime_offset
# Log scaling for frequency values (y-axis)
ax.set_yscale('log')
# Plot times
ax.pcolormesh(time, freq, specgram, norm=norm, cmap=plt.cm.jet)
else:
# this method is much much faster!
specgram = np.flipud(specgram)
# center bin
extent = (time[0] - halfbin_time + xtime_offset,
time[-1] + halfbin_time + xtime_offset,
freq[0] - halfbin_freq, freq[-1] + halfbin_freq)
ax.imshow(specgram,
interpolation="nearest",
extent=extent,
cmap=plt.cm.jet)
if y_lim:
ax.vlines(fig_mark, y_lim[0], y_lim[1], color='r')
else:
ax.vlines(fig_mark, y_lim[0], y_lim[1], color='r')
ax.axis('tight')
# ax.set_xlim(0,end)
# if xtime_offset:
# ax.set_xlim(0+xtime_offset, end+xtime_offset)
# ih = 0
# while 0.5*ih<end:
# ih = ih+1
# xraw_ticks = np.linspace(0,ih*0.5,ih,endpoint=False)
# xnew_ticks = xraw_ticks + xtime_offset
# ax.set_xticks(list(xraw_ticks),list(xnew_ticks))
# else:
# ax.set_xlim(0, end)
if y_lim:
ax.set_ylim(y_lim[0], y_lim[1])
ax.grid(False)
# ax.set_xlabel('Time [s]')
# ax.set_ylabel('Frequency [Hz]')
if title:
ax.set_title(title)
return specgram, freq, time
def plots(self, num_channels=8):
"""
Plot the raw and filtered data
Input:
- num_channels: number of channels to plot
"""
fig = plt.figure()
for channel in range(num_channels):
#fig = plt.figure()
#fig.suptitle(self.name_channel[channel])
self.t_sec = np.array(range(
0, self.raw_eeg_data[:, channel].size)) / self.sample_rate
ax1 = plt.subplot(321)
plt.plot(self.t_sec,
self.raw_eeg_data[:, channel],
label=self.name_channel[channel])
plt.ylabel('EEG (uV)')
plt.xlabel('Time (sec)')
plt.title('Raw')
plt.xlim(self.t_sec[0], self.t_sec[-1])
psd, freqs = mlab.psd(np.squeeze(self.raw_eeg_data[:, channel]),
NFFT=500,
Fs=250)
ax2 = plt.subplot(322)
plt.xlim(self.t_sec[0], self.t_sec[-1])
plt.xlabel('Frequency (Hz)')
plt.ylabel('Power')
plt.ylim(0, 20 * np.median(psd))
plt.plot(freqs, psd, label=self.name_channel[channel])
plt.title('PSD of Unfiltered')
raw_spec_PSDperHz, raw_freqs, raw_t_spec = mlab.specgram(
np.squeeze(raw_eeg_data[:, channel]),
NFFT=NFFT,
window=mlab.window_hanning,
Fs=sample_rate,
noverlap=overlap)
raw_spec_PSDperBin = raw_spec_PSDperHz * sample_rate / float(NFFT)
ax2 = plt.subplot(323)
plt.pcolor(raw_t_spec, raw_freqs,
10 * np.log10(raw_spec_PSDperBin))
plt.clim(25 - 5 + np.array([-40, 0]))
plt.xlim(t_sec[0], t_sec[-1])
plt.ylim([0, 100])
plt.xlabel('Time (sec)')
plt.ylabel('Frequency (Hz)')
plt.title('Spectogram of Unfiltered')
ax3 = plt.subplot(323)
plt.plot(self.t_sec,
self.corrected_eeg_data[:, channel],
label=self.name_channel[channel])
plt.ylabel('EEG (uV)')
plt.xlabel('Time (sec)')
plt.title('Filtered')
plt.xlim(self.t_sec[0], self.t_sec[-1])
psd, freqs = mlab.psd(np.squeeze(self.corrected_eeg_data[:,
channel]),
NFFT=500,
Fs=250)
ax4 = plt.subplot(324)
plt.xlim(self.t_sec[0], self.t_sec[-1])
plt.xlabel('Frequency (Hz)')
plt.ylabel('Power')
plt.ylim(0, 20 * np.median(psd))
plt.plot(freqs, psd, label=self.name_channel[channel])
plt.title('PSD of filtered')
plt.legend(self.name_channel,
loc='upper right',
bbox_to_anchor=(0.01, 0.01))
plt.tight_layout()
plt.show()
def test_specgram_warn_only1seg(self):
"""Warning should be raised if len(x) <= NFFT."""
with pytest.warns(UserWarning, match="Only one segment is calculated"):
mlab.specgram(x=self.y, NFFT=len(self.y), Fs=self.Fs)
