def analyse(filename, resample_to=2756, bt_hop_length=128,
chroma_hop_length=512, chroma_n_fft=1024):
samples, sampleRate = librosa.load(filename)
length = float(len(samples))/sampleRate
if resample_to:
samples = librosa.resample(samples, sampleRate, resample_to)
sampleRate = resample_to
newSampleRate = 2756
samples = librosa.resample(samples, sampleRate, newSampleRate)
sampleRate = newSampleRate
tempo, beats = librosa.beat.beat_track(samples, sampleRate,
hop_length=bt_hop_length)
beat_times = librosa.frames_to_time(beats, sampleRate,
hop_length=bt_hop_length)
chromagram = librosa.feature.chromagram(samples, sampleRate,
hop_length=chroma_hop_length,
n_fft=chroma_n_fft)
chromagram = numpy.transpose(chromagram)
distances = scipy.spatial.distance.cdist(chromagram, CHORDS, "cosine")
chords = distances.argmin(axis=1)
chords = scipy.signal.medfilt(chords, 11)
chord_frames = numpy.array(numpy.where(numpy.diff(chords) != 0))
chords = chords[chord_frames][0].astype(int)
chord_times = librosa.frames_to_time(chord_frames, sampleRate,
hop_length=chroma_hop_length,
n_fft=chroma_n_fft)[0]
chord_names = CHORD_NAMES[chords]
return {"beats": list(beat_times),
"chords": [{"chord": chord_name, "time": chord_time} for chord_name, chord_time in zip(chord_names, chord_times)],
"tempo": tempo}
def test_resample_scikitsamplerate():
warnings.resetwarnings()
warnings.simplefilter('always')
with warnings.catch_warnings(record=True) as out:
librosa.resample(np.zeros(1000), 1000, 500, res_type='sinc_best')
assert len(out) > 0
assert out[0].category is DeprecationWarning
assert 'deprecated' in str(out[0].message).lower()
def lpc_formants(signal, sr, num_formants, max_freq, time_step,
win_len, window_shape='gaussian'):
output = {}
new_sr = 2 * max_freq
alpha = np.exp(-2 * np.pi * 50 * (1 / new_sr))
proc = lfilter([1., -alpha], 1, signal)
if sr > new_sr:
proc = librosa.resample(proc, sr, new_sr)
nperseg = int(win_len * new_sr)
nperstep = int(time_step * new_sr)
if window_shape == 'gaussian':
window = gaussian(nperseg + 2, 0.45 * (nperseg - 1) / 2)[1:nperseg + 1]
else:
window = np.hanning(nperseg + 2)[1:nperseg + 1]
indices = np.arange(int(nperseg / 2), proc.shape[0] - int(nperseg / 2) + 1, nperstep)
num_frames = len(indices)
for i in range(num_frames):
if nperseg % 2 != 0:
X = proc[indices[i] - int(nperseg / 2):indices[i] + int(nperseg / 2) + 1]
else:
X = proc[indices[i] - int(nperseg / 2):indices[i] + int(nperseg / 2)]
frqs, bw = process_frame(X, window, num_formants, new_sr)
formants = []
for j, f in enumerate(frqs):
if f < 50:
continue
if f > max_freq - 50:
continue
formants.append((np.asscalar(f), np.asscalar(bw[j])))
missing = num_formants - len(formants)
if missing:
formants += [(None, None)] * missing
output[indices[i] / new_sr] = formants
return output
def read_song(source_id):
song_data_raw, source_sr = lr.load(data_source[source_id])
song_data = lr.resample(song_data_raw, source_sr, target_sr, scale=True)
song_data = song_data[:song_data.shape[0]/10]
song_data, data_denom = norm(song_data)
return song_data, source_sr, data_denom
def get_best_fs_ratio(a, b, max_drift, steps, max_offset, correlation_size, center=1):
'''
Given two signals with components in common, tries to estimate the clock drift and offset of b vs a
:parameters:
- a : np.ndarray
Some signal
- b : np.ndarray
Some other signal
- max_drift : float
max sample rate drift, in percent, e.g. .02 = 2% clock drift
- steps : int
Number of sample rates to consider, between -max_drift and max_drift
- max_offset : int
Maximum expected offset of the signals
- correlation_size : int
Number of samples to use in each correlate
- center : float
Ratio to deviate from - default 1
Output:
fs_ratio - fs ratio to make b line up well with a
'''
# Sample rate ratios to try
fs_ratios = center + np.linspace(-max_drift, max_drift, steps + 1)
# The max correlation value for each fs ratio
corr_max = np.zeros(fs_ratios.shape)
for n, ratio in enumerate(fs_ratios):
# Resample b with this fs ratio
b_resampled = librosa.resample(b, 1, ratio)
# Compute the max correlation
_, corr = align_over_window(a, b_resampled, max_offset, correlation_size)
corr_max[n] = corr.max()
# Choose ratio with the highest correlation value
return fs_ratios[np.argmax(corr_max)]
def read_song(source_id, target_sr):
song_data_raw, source_sr = lr.load(data_source[source_id])
song_data = lr.resample(song_data_raw, source_sr, target_sr, scale=True)
song_data = song_data[1500:(1500+2*seq_size)] #song_data.shape[0]/10]
song_data, data_denom = norm(song_data, return_denom=True)
return song_data, source_sr, data_denom
def save_waveform_as(self, waveform, data_id, dst):
source_sr, data_denom = self.get_data_info(data_id)
waveform *= data_denom
waveform_resampled = lr.resample(waveform, self.cfg.target_sr, source_sr, scale=True)
logging.info("Saving waveform as {}".format(dst))
lr.output.write_wav(dst, waveform_resampled, source_sr)
def compute_pcen(audio, sr):
# Load settings.
pcen_settings = get_pcen_settings()
# Map to the range [-2**31, 2**31[
audio = (audio * (2**31)).astype('float32')
# Resample to 22,050 kHz
if not sr == pcen_settings["sr"]:
audio = librosa.resample(audio, sr, pcen_settings["sr"])
sr = pcen_settings["sr"]
# Compute Short-Term Fourier Transform (STFT).
stft = librosa.stft(
audio,
n_fft=pcen_settings["n_fft"],
win_length=pcen_settings["win_length"],
hop_length=pcen_settings["hop_length"],
window=pcen_settings["window"])
# Compute squared magnitude coefficients.
abs2_stft = (stft.real*stft.real) + (stft.imag*stft.imag)
# Gather frequency bins according to the Mel scale.
# NB: as of librosa v0.6.2, melspectrogram is type-instable and thus
# returns 64-bit output even with a 32-bit input. Therefore, we need
# to convert PCEN to single precision eventually. This might not be
# necessary in the future, if the whole PCEN pipeline is kept type-stable.
melspec = librosa.feature.melspectrogram(
y=None,
S=abs2_stft,
sr=pcen_settings["sr"],
n_fft=pcen_settings["n_fft"],
n_mels=pcen_settings["n_mels"],
htk=True,
fmin=pcen_settings["fmin"],
fmax=pcen_settings["fmax"])
# Compute PCEN.
pcen = librosa.pcen(
melspec,
sr=pcen_settings["sr"],
hop_length=pcen_settings["hop_length"],
gain=pcen_settings["pcen_norm_exponent"],
bias=pcen_settings["pcen_delta"],
power=pcen_settings["pcen_power"],
time_constant=pcen_settings["pcen_time_constant"])
# Convert to single floating-point precision.
pcen = pcen.astype('float32')
# Truncate spectrum to range 2-10 kHz.
pcen = pcen[:pcen_settings["top_freq_id"], :]
# Return.
return pcen
def slice_clip(filename, start, stop, n_samples, sr, mono=True):
'''Slice a fragment of audio from a file.
This uses pysoundfile to efficiently seek without
loading the entire stream.
Parameters
----------
filename : str
Path to the input file
start : int
The sample index of `filename` at which the audio fragment should start
stop : int
The sample index of `filename` at which the audio fragment should stop (e.g. y = audio[start:stop])
n_samples : int > 0
The number of samples to load
sr : int > 0
The target sampling rate
mono : bool
Ensure monophonic audio
Returns
-------
y : np.ndarray [shape=(n_samples,)]
A fragment of audio sampled from `filename`
Raises
------
ValueError
If the source file is shorter than the requested length
'''
with psf.SoundFile(str(filename), mode='r') as soundf:
n_target = stop - start
soundf.seek(start)
y = soundf.read(n_target).T
if mono:
y = librosa.to_mono(y)
# Resample to initial sr
y = librosa.resample(y, soundf.samplerate, sr)
# Clip to the target length exactly
y = librosa.util.fix_length(y, n_samples)
return y
def _read_song(self, fname, proportion=None):
logging.info("Reading {}".format(fname))
song_data_raw, source_sr = lr.load(fname)
logging.info("Got sampling rate {}, resampling to {} ...".format(source_sr, self.cfg.target_sr))
song_data = lr.resample(song_data_raw, source_sr, self.cfg.target_sr, scale=True)
logging.info("Normalizing with l2 norm ...")
if proportion:
song_data = song_data[: int(proportion * len(song_data)),]
song_data, data_denom = norm(song_data)
logging.info("Done")
return song_data, source_sr, data_denom
def sample_clip(filename, n_samples, sr, mono=True):
'''Sample a fragment of audio from a file.
This uses pysoundfile to efficiently seek without
loading the entire stream.
Parameters
----------
filename : str
Path to the input file
n_samples : int > 0
The number of samples to load
sr : int > 0
The target sampling rate
mono : bool
Ensure monophonic audio
Returns
-------
y : np.ndarray [shape=(n_samples,)]
A fragment of audio sampled randomly from `filename`
Raises
------
ValueError
If the source file is shorter than the requested length
'''
with psf.SoundFile(str(filename), mode='r') as soundf:
n_target = int(np.ceil(n_samples * soundf.samplerate / sr))
# Draw a random clip
start = np.random.randint(0, len(soundf) - n_target)
soundf.seek(start)
y = soundf.read(n_target).T
if mono:
y = librosa.to_mono(y)
# Resample to initial sr
y = librosa.resample(y, soundf.samplerate, sr)
# Clip to the target length exactly
y = librosa.util.fix_length(y, n_samples)
return y
def activation_upsample(A_low, sr):
n, k, _, sr_old = A_low.shape
A = np.zeros((n, k, 1, sr))
for i in range(n):
for j in range(k):
act_res = librosa.resample(A_low[i, j, :, :].squeeze(), orig_sr=sr_old, target_sr=sr)
# Local-max filter act_res
A[i, j, :, : min(sr, len(act_res))] = librosa.localmax(act_res) * act_res
return A
def __test(sr_in, sr_out, res_type, y):
y2 = librosa.resample(y, sr_in, sr_out,
res_type=res_type,
scale=True)
# First, check that the audio is valid
librosa.util.valid_audio(y2, mono=True)
n_orig = np.sqrt(np.sum(np.abs(y)**2))
n_res = np.sqrt(np.sum(np.abs(y2)**2))
# If it's a no-op, make sure the signal is untouched
assert np.allclose(n_orig, n_res, atol=1e-2), (n_orig, n_res)
def analyse(filename, resample_to=2756, bt_hop_length=128, chroma_hop_length=512, chroma_n_fft=1024):
#load audiofile-> return as floating point time series (array) with a sampling rate (int>0)
samples, sampleRate = librosa.load(filename)
length = float(len(samples))/sampleRate
if resample_to:
#resample time series from sampleRate to resample_to rate
samples = librosa.resample(samples, sampleRate, resample_to)
sampleRate = resample_to
newSampleRate = 2756
#resample time series from sampleRate to newSampleRate(2756)
samples = librosa.resample(samples, sampleRate, newSampleRate)
sampleRate = newSampleRate
#track the beats from the time series in samples, using the number of audio samples reps by hop_length
# --> returning tempo(estimated global tempo in bpmin) and beats (frame numbers of estimated beat events)
tempo, beats = librosa.beat.beat_track(samples, sampleRate, hop_length=bt_hop_length)
#convert the frame counts in 'beats' to time (seconds)
beat_times = librosa.frames_to_time(beats, sampleRate, hop_length=bt_hop_length)
#draw a chromagram using the data from samples, sampleRate, hop_length, and window size specified by n_fft
chromagram = librosa.feature.chromagram(samples, sampleRate, hop_length=chroma_hop_length, n_fft=chroma_n_fft)
#permute the dimensions of chromogram into a new array
chromagram = numpy.transpose(chromagram)
#compute the cosine distance between chromogram and CHORDS
distances = scipy.spatial.distance.cdist(chromagram, CHORDS, "cosine")
#return the indices of the minimum values along axis=1 (axis=0 reps the columns, axis=1 reps the rows)
chords = distances.argmin(axis=1)
#apply median filter to chords, the size of the median filter is rep by the kernal of size 11
chords = scipy.signal.medfilt(chords, 11)
#create a new array from th chords array using points where the discrete difference along the x-axis is not 0
chord_frames = numpy.array(numpy.where(numpy.diff(chords) != 0))
chords = chords[chord_frames][0].astype(int)
#convert the frame counts in 'chord_frames' to time(seconds)
chord_times = librosa.frames_to_time(chord_frames, sampleRate, hop_length=chroma_hop_length, n_fft=chroma_n_fft)[0]
chord_names = CHORD_NAMES[chords]
return {"beats": list(beat_times),
"chords": [{"chord": chord_name, "time": chord_time} for chord_name, chord_time in zip(chord_names, chord_times)],
"tempo": tempo}
def __init__(self, file_path=None, raw_samples=None, convert_to_mono=False,
sample_rate=44100, analysis_sample_rate=22050):
"""
Audio constructor.
Opens a file path, loads the audio with librosa, and prepares the features
Parameters
----------
file_path: string
path to the audio file to load
raw_samples: np.array
samples to use for audio output
convert_to_mono: boolean
(optional) converts the file to mono on loading
sample_rate: number > 0 [scalar]
(optional) sample rate to pass to librosa.
Returns
------
An Audio object
"""
if file_path:
y, sr = librosa.load(file_path, mono=convert_to_mono, sr=sample_rate)
elif raw_samples is not None:
# This assumes that we're passing in raw_samples
# directly from another Audio's raw_samples.
y = raw_samples
sr = sample_rate
self.file_path = file_path
self.sample_rate = float(sr)
self.analysis_sample_rate = float(analysis_sample_rate)
self.num_channels = y.ndim
self.duration = librosa.get_duration(y=y, sr=sr)
self.analysis_samples = librosa.resample(librosa.to_mono(y),
sr, self.analysis_sample_rate,
res_type='kaiser_best')
self.raw_samples = np.atleast_2d(y)
self.zero_indexes = self._create_zero_indexes()
self.features = self._create_features()
self.timings = self._create_timings()
def __test(res_type):
y_native, sr = librosa.load(librosa.util.example_audio_file(),
sr=None,
offset=offset,
duration=duration,
res_type=res_type)
y2 = librosa.resample(y_native, sr, sr_target, res_type=res_type)
y, _ = librosa.load(librosa.util.example_audio_file(),
sr=sr_target,
offset=offset,
duration=duration,
res_type=res_type)
assert np.allclose(y2, y)
def signal_to_formants(signal, sr, freq_lims, win_len,
time_step, num_formants, window_shape = 'gaussian',
begin = None, padding = None):
rep = {}
new_sr = 2 * freq_lims[1]
alpha = np.exp(-2 * np.pi * 50 * (1 / new_sr))
proc = lfilter([1., -alpha], 1, signal)
proc = librosa.resample(proc, sr, new_sr)
nperseg = int(win_len*new_sr)
nperstep = int(time_step*new_sr)
if window_shape == 'gaussian':
window = gaussian(nperseg + 2, 0.45 * (nperseg - 1) / 2)[1:nperseg + 1]
else:
window = hanning(nperseg + 2)[1:nperseg+1]
indices = np.arange(int(nperseg / 2), proc.shape[0] - int(nperseg / 2) + 1, nperstep)
num_frames = len(indices)
for i in range(num_frames):
if nperseg % 2 != 0:
X = proc[indices[i] - int(nperseg / 2):indices[i] + int(nperseg / 2) + 1]
else:
X = proc[indices[i] - int(nperseg / 2):indices[i] + int(nperseg / 2)]
frqs, bw = process_frame(X, window, num_formants, new_sr)
formants = []
for j,f in enumerate(frqs):
if f < 50:
continue
if f > freq_lims[1] - 50:
continue
formants.append((np.asscalar(f), np.asscalar(bw[j])))
missing = num_formants - len(formants)
if missing:
formants += [(None,None)] * missing
rep[indices[i]/new_sr] = formants
duration = signal.shape[0] / sr
if begin is not None:
if padding is not None:
begin -= padding
real_output = {}
for k,v in rep.items():
if padding is not None and (k < padding or k > duration - padding):
continue
t = np.asscalar(k+begin)
real_output[t] = v
return real_output
return rep
def get_enrate(signal = None, sr = None, filepath = None, downsample=100):
"""Computes the enrate as described in
ARGS
signal: audio signal <number array>
sr: sampling rate <int>
filepath: fullpath of audio file <str>
downsample: sampling rate to downsample signal <int>
RETURN
enrate: proportional to speaking rate <float>
"""
if signal == None:
signal, sr = load_signal(filepath)
# FFT data
n_fft = downsample
hop_length = int(0.8*downsample)
# Half-wave rectify the signal waveform
signal[signal<0] = 0
# Low-pass filter
numtaps=2
cutoff=16.0
nyq=sr/2.0
transfer_function = scipy.signal.firwin(numtaps=numtaps, cutoff=cutoff/nyq, nyq=nyq)
signal = scipy.signal.lfilter(transfer_function, 1.0, signal)
# Downsample to 100hz
signal = librosa.resample(signal, sr, downsample)
# Hamming window 1-2 seconds with > 75% overlap
fft_window = scipy.signal.hamming(downsample, sym=False)
# FFT, ignore values above 16 hz
magnitudes = np.abs(librosa.stft(signal, n_fft, hop_length, window=fft_window))
bin_count, freq_res = get_bin_count_and_frequency_resolution(n_fft, downsample)
lowest_fbin_idx = int(1/freq_res)
highest_fbin_idx = int(16/freq_res)
# Compute the spectral moment ( index weight each power spectral value and sum )
enrate = np.sum(magnitudes[lowest_fbin_idx:highest_fbin_idx].T * np.array(range(lowest_fbin_idx, highest_fbin_idx)))
return enrate
def wav_data_to_samples(wav_data, sample_rate):
"""Read PCM-formatted WAV data and return a NumPy array of samples.
Uses scipy to read and librosa to process WAV data. Audio will be converted to
mono if necessary.
Args:
wav_data: WAV audio data to read.
sample_rate: The number of samples per second at which the audio will be
returned. Resampling will be performed if necessary.
Returns:
A numpy array of audio samples, single-channel (mono) and sampled at the
specified rate, in float32 format.
Raises:
AudioIOReadError: If scipy is unable to read the WAV data.
AudioIOError: If audio processing fails.
"""
try:
# Read the wav file, converting sample rate & number of channels.
native_sr, y = scipy.io.wavfile.read(six.BytesIO(wav_data))
except Exception as e: # pylint: disable=broad-except
raise AudioIOReadError(e)
if y.dtype == np.int16:
# Convert to float32.
y = int16_samples_to_float32(y)
elif y.dtype == np.float32:
# Already float32.
pass
else:
raise AudioIOError(
'WAV file not 16-bit or 32-bit float PCM, unsupported')
try:
# Convert to mono and the desired sample rate.
if y.ndim == 2 and y.shape[1] == 2:
y = y.T
y = librosa.to_mono(y)
if native_sr != sample_rate:
y = librosa.resample(y, native_sr, sample_rate)
except Exception as e: # pylint: disable=broad-except
raise AudioIOError(e)
return y
def __test(y, sr_in, sr_out, res_type, fix):
y2 = librosa.resample(y, sr_in, sr_out,
res_type=res_type,
fix=fix)
# First, check that the audio is valid
librosa.util.valid_audio(y2, mono=True)
# If it's a no-op, make sure the signal is untouched
if sr_out == sr_in:
assert np.allclose(y, y2)
# Check buffer contiguity
assert y2.flags['C_CONTIGUOUS']
# Check that we're within one sample of the target length
target_length = y.shape[-1] * sr_out // sr_in
assert np.abs(y2.shape[-1] - target_length) <= 1
def __test(infile):
DATA = load(infile)
# load the wav file
(y_in, sr_in) = librosa.load(DATA['wavfile'][0], sr=None, mono=True)
# Resample it to the target rate
y_out = librosa.resample(y_in, DATA['sr_in'], DATA['sr_out'])
# Are we the same length?
if len(y_out) == len(DATA['y_out']):
# Is the data close?
assert np.allclose(y_out, DATA['y_out'])
elif len(y_out) == len(DATA['y_out']) - 1:
assert (np.allclose(y_out, DATA['y_out'][:-1,0]) or
np.allclose(y_out, DATA['y_out'][1:,0]))
elif len(y_out) == len(DATA['y_out']) + 1:
assert (np.allclose(y_out[1:], DATA['y_out']) or
np.allclose(y_out[:-2], DATA['y_out']))
else:
assert False
pass
def resample_all():
audio_folder = 'scenes_stereo/'
subsamp_folder = 'scenes_mono_8k/'
chdir(audio_folder)
mkdir(subsamp_folder)
for sub_folder in glob('*'):
mkdir(subsamp_folder + sub_folder)
for filename in glob(sub_folder + '/*.wav'):
print(filename)
[fs, sig] = read(filename)
sig = to_mono(sig.T)
sig = resample(sig, fs, 8000)
write(subsamp_folder + filename, 8000, sig)
def apply_offsets_resample(b, offset_locations, offsets):
'''
Adjust a signal b according to local offset estimations using resampling
:parameters:
- b : np.ndarray
Some signal
- offset_locations : np.ndarray
locations, in samples, of each local offset estimation
- offsets : np.ndarray
local offset for the corresponding sample in offset_locations
:returns:
- b_aligned : np.ndarray
b with offsets applied
'''
assert offset_locations.shape[0] == offsets.shape[0]
# Include signal boundaries in offset locations
offset_locations = np.append(0, np.append( offset_locations, b.shape[0]-100 ))
# Allocate output signal
b_aligned = np.zeros(np.int(np.sum(np.diff(offset_locations)) + np.max(np.abs(offsets))))
# Set last offset to whatever the second to last one was
offsets = np.append(offsets, offsets[-1])
current = 0
# !!!!!!!!!!!!!!!!!!
# Should zip here
# !!!!!!!!!!!!!!!!!!
for n, offset in enumerate(offsets):
start = offset_locations[n]
end = offset_locations[n + 1]
# Compute the necessary resampling ratio to compensate for this offset
ratio = 1 + (-offset + start - current)/(end - start)
# Resample this portion of the signal, with some padding at the end
resampled = librosa.resample(b[start:end + 100], 1, ratio)
# Compute length and place the signal
length = int(end - current - offset)
b_aligned[current:current + length] = resampled[:length]
current += length
return b_aligned
def __test(infile, scipy_resample):
DATA = load(infile)
# load the wav file
(y_in, sr_in) = librosa.load(DATA['wavfile'][0], sr=None, mono=True)
# Resample it to the target rate
y_out = librosa.resample(y_in, DATA['sr_in'], DATA['sr_out'],
scipy_resample=scipy_resample)
# Are we the same length?
if len(y_out) == len(DATA['y_out']):
# Is the data close?
assert np.allclose(y_out, DATA['y_out'])
elif len(y_out) == len(DATA['y_out']) - 1:
assert (np.allclose(y_out, DATA['y_out'][:-1, 0]) or
np.allclose(y_out, DATA['y_out'][1:, 0]))
elif len(y_out) == len(DATA['y_out']) + 1:
assert (np.allclose(y_out[1:], DATA['y_out']) or
np.allclose(y_out[:-2], DATA['y_out']))
else:
assert False
pass
def read_csv(self):
waves = []
labels = {}
with open(self.dataset_path, "r") as file:
csv_reader = csv.reader(file)
for row in csv_reader:
if row:
label = row[-1]
wav_path = f"{CURRENT_DIR}/{row[0]}" if row[0][
0] != '/' else f"{CURRENT_DIR}{row[0]}"
signal, sr = librosa.load(wav_path, sr=self.dim)
signal = librosa.resample(signal, sr, self.dim)
if signal.shape[0] != self.dim:
continue
waves.append(wav_path)
labels[wav_path] = label
return waves, labels
def resample(y, src_sr, target_sr, mode='kaiser_fast'):
if mode == 'kaiser_best':
warnings.warn(
f'Using resampy in kaiser_best to {src_sr}=>{target_sr}. This function is pretty slow, \
we recommend the mode kaiser_fast in large scale audio trainning')
assert type(y) == np.ndarray, 'currently only numpy data are supported'
assert mode in __resample_mode__, f'resample mode must in {__resample_mode__}'
assert type(
src_sr
) == int and src_sr > 0 and src_sr <= 48000, 'make sure type(sr) == int and sr > 0 and sr <= 48000,'
assert type(
target_sr
) == int and target_sr > 0 and target_sr <= 48000, 'make sure type(sr) == int and sr > 0 and sr <= 48000,'
if has_resampy:
return resampy.resample(y, src_sr, target_sr, filter=mode)
if has_librosa:
return librosa.resample(y, src_sr, target_sr, res_type=mode)
assert False, 'requires librosa or resampy to do resampling, pip install resampy'
def on_post(self, request, response):
# NB: uses middleware to pull out data.
form_data = request.params['audio_data'].file
data, samplerate = soundfile.read(form_data)
# For debugging browser input, uncomment the following line:
# scipy.io.wavfile.write('browser_input_audio.wav', samplerate, data)
# NB: Convert the input stereo signal into mono.
# In the future the frontend should be responsible for sampling details.
mono = data[:, 0]
# NB: We must downsample to the rate that the network is trained on.
downsampled = librosa.resample(mono, samplerate, 16000)
# Evaluate the model
print(">>> Converting...")
results = converter.convert(downsampled, conversion_direction = 'A2B')
temp_dir = tempfile.TemporaryDirectory(prefix='tmp_ml_audio')
temp_file = tempfile.NamedTemporaryFile(suffix='.wav')
temp_file.write(results.read())
out_file = temp_dir.name + '/output.ogg'
# NB: Browsers have a great deal of trouble decoding WAV files unless they are in the
# narrow slice of the WAV spec expected. None of the {librosa, scipy, soundfile} python
# tools do a good job of this, so here we shell out to ffmpeg and generate OGG.
# It's lazy and messy, but it works for now.
# See https://github.com/librosa/librosa/issues/361 for a survey of the library landscape
# See https://bugzilla.mozilla.org/show_bug.cgi?id=523837 for one of dozens of browser codec bugs
_stdout = subprocess.check_output(['ffmpeg', '-i', temp_file.name, '-acodec', 'libvorbis', out_file])
response.content_type = 'audio/ogg'
with open(out_file, mode='rb') as f:
response.data = f.read()
def resample2(data, sr):
'''Resample if required and drop to disc resampled data '''
# now = datetime.datetime.now()
# time = now.strftime("%H:%M:%S")
# print(time, ': Resample all recordings to target sampling rate if required')
downsample = data.loc[data['sample_rate'] != sr]
#print(file, ': Resampling records:', len(downsample.index))
resampled = []
indexes = []
length = []
srs = []
samples = []
for index, row in downsample.iterrows():
sample = row['raw_sounds']
sro = row['sample_rate']
y = librosa.resample(sample, sro, sr)
l = len(y) / sr
resampled.append(y)
indexes.append(index)
length.append(l)
samples.append(len(y))
srs.append(sr)
output = pd.DataFrame(list(zip(resampled, srs, indexes, length, samples)),
columns=[
'raw_sounds', 'sample_rate', 'index', 'length',
'sample_count'
]).set_index('index')
data.update(output) # Join resampled recordings to raw frame
# data.to_pickle(p / file.name)
#file_list = [p/(file.name) for file in file_list]
return data
def process_uid(uid):
audio_path = os.path.join(audio_files_dir, uid)
sound_class = uid.split('-')[-1].split('.wav')[0]
if sound_class not in sound_classes:
raise ValueError('Sound Class: {} must be in Classes: {}'.format(
sound_class, sound_classes))
y, sr = librosa.load(audio_path, sr=44100)
y_8k = librosa.resample(y, sr, 8000)
if y_8k.shape[0] < min_wav_samples:
return
wav = torch.tensor(y_8k, dtype=torch.float32).unsqueeze(0)
norm_wav = torch.tensor(normalize_wav(y_8k),
dtype=torch.float32).unsqueeze(0)
output_uid_folder = os.path.join(output_dirpath, sound_class,
uid.split('.wav')[0])
data = {
'wav': wav,
'wav_norm': norm_wav,
}
# Append metadata to the written data
for meta_label, meta_label_val in metadata_dict[uid].items():
if meta_label in data:
raise IndexError('Trying to override essential '
'information about files by '
'assigning metalabel: {} in data '
'dictionary: {}'.format(meta_label, data))
data[meta_label] = meta_label_val
if not os.path.exists(output_uid_folder):
os.makedirs(output_uid_folder)
for k, v in data.items():
file_path = os.path.join(output_uid_folder, k)
joblib.dump(v, file_path, compress=0)
def load_audio(self):
"""
Reads wav file based on csv values, resamples audio to 8000hz, fixes length to 1 second
:return: numpy array of stereo audio, DOA from file
"""
df = pd.read_csv("{dir}/iteration_{iter}.csv".format(dir=self.directory, iter=self.iteration),
usecols=[1, 2])
wav_name = df.iloc[0][0]
filename = "{wav_name}".format(wav_name=wav_name)
y, sr = librosa.load(filename, mono=False)
y_8k = librosa.resample(y, sr, 8000)
o_env = librosa.onset.onset_strength(y_8k[0], sr=8000)
peaks = librosa.util.peak_pick(o_env, 3, 3, 3, 5, 0.25, 5)
times = librosa.frames_to_time(np.arange(len(o_env)),
sr=8000, hop_length=512)
peak_times = times[peaks]
time = 0
for i in range(1, len(peak_times) + 1):
if 3 - peak_times[-i] >= 0.75:
time = peak_times[-i] - 0.25
break
sample = librosa.time_to_samples(np.array([time]), sr=8000)
sliced_y = np.array([y_8k[0][sample[0]:], y_8k[1][sample[0]:]])
y_out = librosa.util.fix_length(sliced_y, 8000)
return y_out
def import_to_mel(filepath, sample_rate):
'''
Import target audio file and pre process.
input:
filepath to target audio
sample_rate of the current model
output:
a mel spectrogram of the loaded audiofile, normalised and limited to 4 seconds.
'''
# mel settings
n_mels = 128
n_fft = 2048
hop_length = 512 # the paper says 2048, but then the output matrix is the wrong size ЪциРђЇРЎѓ№ИЈ
# import, convert to mono, normalise
waveform, sr = torchaudio.load(filepath)
waveform = waveform.numpy()
if (sr != sample_rate):
waveform = librosa.resample(waveform, sr, sample_rate)
if (waveform.shape[0] > 1):
waveform = librosa.to_mono(waveform).reshape(1, len(waveform[0]))
waveform[0] = waveform[0] * (1.0 / np.max(waveform[0])) # normalise
waveform = torch.from_numpy(waveform)
# copy to a tensor of specific size
waveform_4s = torch.zeros(1, sample_rate * 4)
iter_len = min(sample_rate * 4, waveform.shape[1])
for i in range(iter_len):
waveform_4s[0][i] = waveform[0][i]
# generate mel
spectrogram = torchaudio.transforms.MelSpectrogram(
sample_rate=sample_rate,
n_mels=n_mels,
n_fft=n_fft,
hop_length=hop_length)(waveform_4s)
return spectrogram
def audio_strip(self, orig_x, orig_sr) -> np.array:
"""
Trim silence from both side
"""
# Noise reduction
if self._do_noise_reduction:
orig_x = nr.reduce_noise(audio_clip=orig_x, noise_clip=orig_x)
# Resampling
x = librosa.resample(orig_x,
orig_sr=orig_sr,
target_sr=self.SAMPLE_RATE)
# Voice activity detection
start_sample = 0
stop_sample = x.shape[0]
# begin
for i in range(x.shape[0] // self.VAD_CHUNK):
frame_buf = self.float_to_pcm16(x[self.VAD_CHUNK *
i:self.VAD_CHUNK * (i + 1)])
if self._vad.is_speech(frame_buf, self.SAMPLE_RATE):
start_sample = (i + 1) * self.VAD_CHUNK
break
# end
for i in range(x.shape[0] // self.VAD_CHUNK - 1,
start_sample // self.VAD_CHUNK, -1):
frame_buf = self.float_to_pcm16(x[self.VAD_CHUNK *
i:self.VAD_CHUNK * (i + 1)])
if self._vad.is_speech(frame_buf, self.SAMPLE_RATE):
stop_sample = i * self.VAD_CHUNK
break
# less than 120 ms
if stop_sample - start_sample <= self.MINIMUM_DIFF * (
self.SAMPLE_RATE // 1000):
return None
return x[start_sample:stop_sample]
def preprocess(raw_path, clean_path):
raw_path = os.path.join(BASE_DIR, raw_path)
clean_path = os.path.join(BASE_DIR, clean_path)
if not os.path.isdir(raw_path):
print('Path to Raw dataset is invalid!')
return
if not os.path.isdir(clean_path):
os.mkdir(clean_path)
for cls in os.scandir(raw_path):
if cls.is_dir:
for item in tqdm(os.scandir(cls), total=len(os.listdir(cls))):
if item.name.endswith('.wav'):
audio = item.path
sr, signal = wavfile.read(audio)
signal = signal.astype(np.float32).T
if signal.shape[0] == 2:
signal = to_mono(signal)
elif signal.shape[0] == 1:
signal = to_mono(signal.reshape(-1))
signal = resample(signal, sr, SR)
sr = SR
signal = signal.astype(np.int16)
mask = purifier(signal, sr, 100)
signal = signal[mask]
if signal.shape[
0] < DURATION: #if the audio after purification is less than 2s, append zeros
rectified_signal = np.zeros((DURATION, ),
dtype=np.int16)
rectified_signal[:signal.shape[0]] = signal
save_file(signal, sr, cls.name, item.name, 0)
else:
trunc = signal.shape[0] % DURATION
for i, j in enumerate(
range(0, signal.shape[0] - trunc, DURATION)):
strip = signal[j:j + DURATION]
save_file(strip, sr, clean_path, cls.name,
item.name, i)
def load_wav_file(file_path,
sample_rate,
mono=True,
resample_type="kaiser_best"):
"""Load a wav audio file as a floating point time series. Significantly faster than
load_sound_file."""
actual_sample_rate, samples = wavfile.read(file_path)
if samples.dtype != np.float32:
assert samples.dtype == np.int16
samples = np.true_divide(
samples, 32768,
dtype=np.float32) # ends up roughly between -1 and 1
if mono and len(samples.shape) > 1:
if samples.shape[1] == 1:
samples = samples[:, 0]
else:
samples = np.mean(samples, axis=1)
if sample_rate is not None and actual_sample_rate != sample_rate:
if resample_type == "auto":
resample_type = ("kaiser_fast" if actual_sample_rate < sample_rate
else "kaiser_best")
samples = librosa.resample(samples,
actual_sample_rate,
sample_rate,
res_type=resample_type)
warnings.warn(
"{} had to be resampled from {} hz to {} hz. This hurt execution time."
.format(str(file_path), actual_sample_rate, sample_rate))
actual_sample_rate = actual_sample_rate if sample_rate is None else sample_rate
return samples, actual_sample_rate
def generator(self,
data_dir,
tmp_dir,
dataset,
eos_list=None,
start_from=0,
how_many=0):
del eos_list
i = 0
data_tuples = _collect_data(tmp_dir)
encoders = self.feature_encoders(data_dir)
audio_encoder = encoders["waveforms"]
text_encoder = encoders["targets"]
for utt_id, media_file, text_data, speaker, utt_dataset in tqdm(
sorted(data_tuples)[start_from:]):
if dataset != utt_dataset:
continue
if how_many > 0 and i == how_many:
return
i += 1
try:
wav_data = audio_encoder.encode(media_file)
except AssertionError:
audio, sr = librosa.load(media_file)
data_resampled = librosa.resample(audio, sr, SAMPLE_RATE)
with tempfile.NamedTemporaryFile(suffix='.wav') as fid:
librosa.output.write_wav(fid.name, data_resampled,
SAMPLE_RATE)
wav_data = audio_encoder.encode(fid.name)
yield {
"waveforms": wav_data,
"waveform_lens": [len(wav_data)],
"targets": text_encoder.encode(text_data),
"raw_transcript": [text_data],
"utt_id": [utt_id],
"spk_id": [speaker],
}
def _get_spectrograms(fpath, require_sr, preemphasis, n_fft, hop_length,
win_length, max_db, ref_db):
'''Parse the wave file in `fpath` and
Returns normalized melspectrogram and linear spectrogram.
Args:
fpath: A string. The full path of a sound file.
Returns:
mel: A 2d array of shape (T, n_mels) and dtype of float32.
mag: A 2d array of shape (T, 1+n_fft/2) and dtype of float32.
'''
# Loading sound file
y, sr = librosa.load(fpath, sr=None)
if sr != require_sr:
y = librosa.resample(y, sr, require_sr)
# Preemphasis
y = np.append(y[0], y[1:] - preemphasis * y[:-1])
# stft
linear = librosa.stft(y=y,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length)
# magnitude spectrogram
mag = np.abs(linear) # (1+n_fft//2, T)
# to decibel
mag = 20 * np.log10(np.maximum(1e-5, mag))
# normalize
mag = np.clip((mag - ref_db + max_db) / max_db, 1e-8, 1)
# Transpose
mag = mag.T.astype(np.float32) # (T, 1+n_fft//2)
return mag
def wav_data_to_samples(wav_data, sample_rate):
"""Read PCM-formatted WAV data and return a NumPy array of samples.
Uses scipy to read and librosa to process WAV data. Audio will be converted to
mono if necessary.
Args:
wav_data: WAV audio data to read.
sample_rate: The number of samples per second at which the audio will be
returned. Resampling will be performed if necessary.
Returns:
A numpy array of audio samples, single-channel (mono) and sampled at the
specified rate, in float32 format.
Raises:
AudioIOReadException: If scipy is unable to read the WAV data.
AudioIOException: If audio processing fails.
"""
try:
# Read the wav file, converting sample rate & number of channels.
native_sr, y = scipy.io.wavfile.read(six.BytesIO(wav_data))
except Exception as e: # pylint: disable=broad-except
raise AudioIOReadException(e)
if y.dtype != np.int16:
raise AudioIOException('WAV file not 16-bit PCM, unsupported')
try:
# Convert to float, mono, and the desired sample rate.
y = int16_samples_to_float32(y)
if y.ndim == 2 and y.shape[1] == 2:
y = y.T
y = librosa.to_mono(y)
if native_sr != sample_rate:
y = librosa.resample(y, native_sr, sample_rate)
except Exception as e: # pylint: disable=broad-except
raise AudioIOException(e)
return y
def asr_transcript(model, tokenizer, input_file):
if not os.path.isfile(input_file):
raise FileNotFoundError
# tokenizer, model = load_model()
speech, fs = sf.read(input_file)
if len(speech.shape) > 1:
speech = speech[:, 0] + speech[:, 1]
if fs != 16000:
speech = librosa.resample(speech, fs, 16000)
input_values = tokenizer(speech, return_tensors="pt").input_values
input_values = input_values.to(device)
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.decode(predicted_ids[0])
return correct_sentence(transcription.lower())
def supercompression(filename, mono, resample):
y_file, fs = librosa.load(filename, mono=mono)
if resample:
targetsample = 2000
y_file = librosa.resample(y_file, target_sr=targetsample, orig_sr=fs)
D = librosa.stft(y_file)
S_db = np.transpose(
np.array(librosa.amplitude_to_db(np.abs(D), ref=np.max))).tolist()
edited = []
average = 0
count = 0
for line in S_db:
temp = []
for data in line:
data += 80
if data > 0:
average += (data)
count += 1
average = average / count
for line in S_db:
temp = []
for data in line:
data += 80
if data == 0:
category = 0
elif data > 0 and data < average * 0.5:
category = 1
elif data > average * 0.5 and data < average:
category = 2
elif data > average and data < average * 1.5:
category = 3
elif data > average * 1.5:
category = 4
temp.append(category)
edited.append(temp)
return edited
def read_audio():
# 녹음을 수행하고 prediction을 수행하는 부분
audio = sd.rec(duration * fs, samplerate=fs, channels=4, dtype='float64')
# print("Recording Audio")
sd.wait()
audio = np.multiply(audio, negative)
audio = np.multiply(audio, mul)
audio = np.multiply(audio, negative)
# sf.write('./test audio file/False_alarm.wav', audio, fs)
# print('recorded!')
audio = audio.T
y = librosa.resample(audio, fs, 16000)
y = np.asfortranarray(y)
fe1, fe2, fe3, fe4 = extract_feature(y, sr=16000)
rr1, rr2 = prediction(fe1, fe2, fe3, fe4)
result = ''
if not rr2:
result = rr1
else:
result = rr1 + ", " + rr2
return result
def __getitem__(self, idx):
hop_length = 1024
# open audio
file_path = self.data[idx]
signal, sampling_rate = open_audio(file_path)
if len(signal.shape) > 1:
signal = np.mean(signal, axis = 1)
if sampling_rate != 44100:
signal = librosa.resample(signal, sampling_rate, 44100)
sampling_rate = 44100
# get 30 second chunk
len_index_30_sec = int(30 / (1 / sampling_rate))
# trim first and last 30 seconds
signal = signal[len_index_30_sec:-len_index_30_sec]
# random start index
start_index = np.random.randint(low = 0, high = len(signal) - len_index_30_sec)
signal = signal[start_index:start_index + len_index_30_sec]
# if training change pitch randomly
if self.train:
n_steps = np.random.randint(low = -4, high=4)
signal = librosa.effects.pitch_shift(signal, sampling_rate, n_steps=n_steps)
# extract harmonic
data_h = librosa.effects.harmonic(signal)
# cqt transform
S = np.real(librosa.cqt(data_h, sr=sampling_rate, hop_length=hop_length)).astype(np.float32)
d = torch.from_numpy(np.expand_dims(S, axis = 0)).type(torch.FloatTensor)
# normalize
d = F.normalize(d)
l = torch.from_numpy(np.array(self.labels[idx])).type(torch.LongTensor)
# print(d.shape, sampling_rate, file_path)
return d,l
def generate_amplitude_envelopes(signal, sr, num_bands, min_frequency, max_frequency, mode='downsample'):
signal = preemphasize(signal, 0.97)
proc = signal / np.sqrt(np.mean(signal ** 2)) * 0.03
band_mins = [min_frequency * np.exp(np.log(max_frequency / min_frequency) / num_bands) ** x
for x in range(num_bands)]
band_maxes = [min_frequency * np.exp(np.log(max_frequency / min_frequency) / num_bands) ** (x + 1)
for x in range(num_bands)]
envs = []
for i in range(num_bands):
b, a = butter(2, (band_mins[i] / (sr / 2), band_maxes[i] / (sr / 2)), btype='bandpass')
env = filtfilt(b, a, proc)
env = abs(hilbert(env))
if mode == 'downsample':
env = resample(env, sr, 120)
envs.append(env)
envs = np.array(envs).T
if mode == 'downsample':
sr = 120
output = dict()
for i in range(envs.shape[0]):
output[i / sr] = envs[i, :]
return output
def preprocess_wav(fpath_or_wav, source_sr = None):
"""
Applies the preprocessing operations used in training the Speaker Encoder to a waveform
either on disk or in memory. The waveform will be resampled to match the data hyperparameters.
:param fpath_or_wav: either a filepath to an audio file (many extensions are supported, not
just .wav), either the waveform as a numpy array of floats.
:param source_sr: if passing an audio waveform, the sampling rate of the waveform before
preprocessing. After preprocessing, the waveform's sampling rate will match the data
hyperparameters. If passing a filepath, the sampling rate will be automatically detected and
this argument will be ignored.
"""
# Load the wav from disk if needed
if isinstance(fpath_or_wav, str) or isinstance(fpath_or_wav, Path):
wav, source_sr = librosa.load(fpath_or_wav, sr=None)
else:
wav = fpath_or_wav
# Resample the wav if needed
if source_sr is not None and source_sr != sampling_rate:
wav = librosa.resample(wav, source_sr, sampling_rate)
## Apply the preprocessing: normalize volume and shorten long silences
#wav = normalize_volume(wav, audio_norm_target_dBFS, increase_only=True)
#wav = trim_long_silences(wav)
return wav
def pool_msr(path, dirname, mrs_dir, msf_dir, ftype = 3):
dirs = os.listdir("%s/%s"%(path, dirname))
for file in dirs:
if file.endswith('.wav'):
print('%s/%s/%s' % (path, dirname, file))
if dirname == "":
sample_rate, signal = scipy.io.wavfile.read(path, file)
feat = logfbank(signal)
mf = srmr_audio(path, file, pad)
else:
signal, fs = librosa.load("%s/%s/%s" % (path, dirname, file))
signal = librosa.resample(signal, fs, 16000)
feat1 = logfbank(signal)
feat2 = ssc(signal)
mrs = feat1
write_file(mrs_dir, file, mrs, ftype)
msf = feat2
write_file(msf_dir, file, msf, ftype)
def preview(filename, directory):
print '...Previewing Accompaniment'
FNULL = open(os.devnull, 'w')
subprocess.call([
'fluidsynth', '-T', 'wav', '-F', directory + "/" + filename[:-4] +
'.raw', '-ni', directory[:-10] + 'lib/sf2/sf.sf2',
directory + "/" + filename[:-4] + '.mid', '-g', '0.8', '-r', '22050'
],
stdout=FNULL,
stderr=subprocess.STDOUT)
subprocess.call([
'SoX', '-t', 'raw', '-r', '22050', '-e', 'signed', '-b', '16', '-c',
'1', directory + "/" + filename[:-4] + '.raw',
directory + "/" + filename[:-4] + '_midi.wav'
])
y, sr = librosa.load(directory + "/" + filename)
z, sr2 = librosa.load(directory + "/" + filename[:-4] + '_midi.wav')
y = librosa.resample(y, sr, sr * 2)
mix = np.zeros(max(len(y), len(z)), dtype=float)
mix[:len(y)] += y / 2
mix[:len(z)] += z / 2
mix = np.int16(mix / np.max(np.abs(mix)) * 16383)
write(directory + "/" + filename[:-4] + '_mix.wav', 44100, mix)
def process_wav_file(wav, orig_d_path, target_d_path, sample_duration,
sample_rate, channels_to_extract):
# read the orig wav file, and resample with given sample_rate
wav_p = os.path.join(orig_d_path, wav)
data, sr = sf.read(wav_p, dtype=np.float32)
if data.shape[1] < channels_to_extract:
raise ValueError('Not enough channels')
# reduce it to a single channel and resample
data = np.concatenate(data[:, :channels_to_extract])
data = librosa.resample(data, sr, sample_rate)
# split it into the required duration
frames_per_sample = int(sample_duration * sample_rate)
# write the audio to the given directory
idx = 0
start = 0
end = frames_per_sample
while end < data.shape[0]:
target_wav = wav[:-4] + '_' + str(idx) + '.wav'
target_wav_p = os.path.join(target_d_path, target_wav)
wavfile.write(target_wav_p, sample_rate, data[start:end])
idx += 1
start = end
end += frames_per_sample
def get_spectrogram_feature(filepath):
if filepath.split('/')[1] == 'TIMIT':
sig = np.fromfile(filepath, dtype=np.int16)[512:].reshape((-1, 1))
else:
(fate, width, sig) = wavio.readwav(filepath)
sig = sig.ravel().astype(np.float) / 32767
sig = librosa.resample(sig, 16000, 8000) * 32767
sig = sig.astype(np.int16)
stft = torch.stft(torch.FloatTensor(sig),
N_FFT,
hop_length=int(0.01 * SAMPLE_RATE),
win_length=int(0.03 * SAMPLE_RATE),
window=torch.hamming_window(int(0.03 * SAMPLE_RATE)),
center=False,
normalized=False,
onesided=True)
stft = (stft[:, :, 0].pow(2) + stft[:, :, 1].pow(2)).pow(0.5)
amag = stft.numpy()
feat = torch.FloatTensor(amag)
feat = torch.FloatTensor(feat).transpose(0, 1)
return feat
def get_x_y(i, fs, resample, directory, input_data_frame):
"""
returns tuple of x, y data pair at sample rate fs
written to ensure that parallel computations have the correct label
"""
file_path = "{directory}/{filename}".format(
directory=directory, filename=input_data_frame.iloc[i][0])
y, sr = librosa.load(file_path, mono=False)
if resample:
y_8k = librosa.resample(y, sr, fs)
result_x = librosa.util.fix_length(y_8k, fs)
else:
y_8k = y
result_x = librosa.util.fix_length(y_8k, sr)
result_y = get_y(i, input_data_frame)
return result_x, result_y
def load_audio(audiofile):
try:
audio, sr = soundfile.read(audiofile)
if audio.shape[1] != 1:
audio = librosa.to_mono(audio.T)
if sr != 16000:
audio = librosa.resample(audio, sr, 16000)
except:
path_audio = Path(audiofile)
filetype = path_audio.suffix
assert filetype in ['.mp3', '.ogg', '.flac', '.wav', '.m4a',
'.mp4'], filetype
with tempfile.TemporaryDirectory() as tempdir:
tempwav = Path(tempdir) / (path_audio.stem + '_temp' + '.flac')
command = [
'ffmpeg', '-i', audiofile, '-af', 'aformat=s16:16000', '-ac',
'1', tempwav
]
process = subprocess.Popen(command,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
stdout, stderr = process.communicate()
audio, sr = soundfile.read(tempwav)
return audio
def spoil_audio(sample, sr=8000, rem_coef=0.1):
"""Spoil the wav audio file by removing measures.
Inputs:
sample (numpy.ndarray): wav file array;
sr=8000 (int): incoming sample rate;
rem_coef=0.1 (float): removing coefficient
Return:
numpy.ndarray: reprocessed sample wav file array"""
new_sample_rate = sr + sr * rem_coef
sample_filtered = librosa.resample(sample,
orig_sr=sr,
target_sr=new_sample_rate)
# Получить Series из wav numpy.ndarray
s1 = Series(sample_filtered)
# Выбрать случайно индексы
rand_ind = np.random.choice(s1.index.values, sr, replace=False)
# Выбрать из исходного Series только полученные случайные индексы
s1.loc[rand_ind].sort_index().values
return s1.loc[rand_ind].sort_index().values
def _resample_and_cut(cache_path, spec, example, label):
"""Resample and cut the audio files into snippets."""
def _new_path(i):
# Note that `splitext` handles hidden files differently and here we just
# assume that audio files are not hidden files.
# os.path.splitext('/root/path/.wav') => ('/root/path/.wav', '')
# instead of ('/root/path/', '.wav') as this code expects.
filename_without_extension = os.path.splitext(
os.path.basename(example))[0]
new_filename = filename_without_extension + '_%d.wav' % i
return os.path.join(cache_path, label, new_filename)
sampling_rate, xs = wavfile.read(example)
if xs.dtype != np.int16:
raise ValueError(
'DataLoader expects 16 bit PCM encoded WAV files, but {} has type {}'
.format(example, xs.dtype))
# Extract snippets.
n_samples_per_snippet = int(spec.snippet_duration_sec * sampling_rate)
begin_index = 0
count = 0
while begin_index + n_samples_per_snippet <= len(xs):
snippet = xs[begin_index:begin_index + n_samples_per_snippet]
if spec.target_sample_rate != sampling_rate:
# Resample, librosa.resample only works with float32.
# Ref: https://github.com/bmcfee/resampy/issues/44
snippet = snippet.astype(np.float32)
snippet = librosa.resample(
snippet,
orig_sr=sampling_rate,
target_sr=spec.target_sample_rate).astype(np.int16)
wavfile.write(_new_path(count), spec.target_sample_rate, xs)
begin_index += n_samples_per_snippet
count += 1
return count
def get_audio(config, mp3_path=None, array=None, array_sr=None):
if mp3_path:
array, sr_in = librosa.core.load(mp3_path, sr=None, mono=False)
elif array is not None:
array = array.astype(np.float32)
sr_in = array_sr
array = librosa.core.to_mono(array)
array = librosa.resample(array, sr_in, config.sr)
array = librosa.core.power_to_db(
librosa.feature.melspectrogram(array, config.sr, n_mels=config.n_mels))
array = array.astype(np.float32)
# normalization
mean, variance = tf.nn.moments(tf.constant(array),
axes=[0, 1],
keepdims=True)
array = tf.nn.batch_normalization(array,
mean,
variance,
offset=0,
scale=1,
variance_epsilon=.000001).numpy()
return array
def __onset_time(song_path):
'''
Loads the song at song_path, computes onsets, returns array of times and harmonic component
of source separated audio
~~~~ ARGUMENTS ~~~~
- song_path (Path or str): path to audio file
~~~~ RETURNS ~~~~
- y_harmonic (1D numpy array): numpy representation of audio, sr=22050, with percussive
components of audio removed
- onset_times (list of float): list of onset times corresponding to audio in seconds
'''
# Load the songs and the notes arrays one at a time
# for idx in range (len(song_paths)):
# Load the song
y, sr = librosa.load(song_path)
# resample the song if it isn't sr=22050 (for consistent sizing)
if not sr == 22050:
y = librosa.resample(y, sr, 22050)
sr = 22050
#source seperation, margin can be tuned
y_harmonic, _ = librosa.effects.hpss(y, margin=2.0)
# Set Hop_len
hop_len = 512
onset_frame_backtrack = librosa.onset.onset_detect(y_harmonic,
sr=sr,
hop_length=hop_len,
backtrack=True)
onset_times = librosa.frames_to_time(onset_frame_backtrack)
return y_harmonic, onset_times
def resample(sample_rate=None, dir=None, csv_path=None):
clips = []
start_time = time.time()
# List all clips that appear on the csv (train, eval or test)
if csv_path != 'test':
with open(csv_path, 'r') as csvFile:
reader = csv.reader(csvFile)
for row in reader:
clips.append(row[0])
csvFile.close()
clips.remove('fname')
else:
clips = os.listdir(dir)
if os.path.exists(dir+'/resampled/'):
shutil.rmtree(dir+'/resampled', ignore_errors=True) # ignore errors whit read only files
os.mkdir(dir+'/resampled')
for clip in clips:
# Audio clip is read
data, sr = sf.read(dir+'/'+clip)
data = data.T
# Audio data is resampled to desired sample_rate
if sr != sample_rate:
data_resampled = librosa.resample(data, sr, sample_rate)
# Processed data is saved into a directory under train_clip_dir
sf.write(dir+'/resampled/'+clip, data_resampled, sample_rate, subtype='PCM_16')
print('Audio data has been resampled successfully')
elapsed_time = time.time() - start_time
print('Elapsed time ' + str(elapsed_time) + ' seconds')
def k_filter(data, fs):
"""
x_t: audio data in samples across time
fs: sample rate of x_t
TEMPORARY FUNCTION UNTIL THE LOUDNESS FUNCTION IS FIXED FOR ACTIVITY
return: k-filtered data AND new 48khz fs
"""
# Convert fs to 48khz to do K-Filtering
if fs != 48000:
data = librosa.resample(data, fs, 48000)
fs = 48000
# Hi-Shelf Boost of +4dB at 1681hz
a1 = [1.0, -1.69065929318241, 0.73248077421585]
b1 = [1.53512485958697, -2.69169618940638, 1.19839281085285]
# Create High-Pass roll off at 38hz
a2 = [1.0, -1.99004745483398, 0.99007225036621]
b2 = [1.0, -2.0, 1.0]
# Filter in succession
return lfilter(b2, a2, lfilter(b1, a1, data)), fs
def analyze_sound(event, context):
file_data = event
file_name = file_data['name']
bucket_name = file_data['bucket']
blob = storage_client.bucket(bucket_name).get_blob(file_name)
blob_name = blob.name
_, temp_local_filename = tempfile.mkstemp()
# Download file from bucket.
blob.download_to_filename(temp_local_filename)
y, sr = librosa.load(temp_local_filename)
y_mono = librosa.to_mono(y)
y_mono_22050 = librosa.resample(y_mono, sr, 22050)
mfccs = librosa.feature.mfcc(y=y_mono_22050, sr=22050, n_mfcc=12)
mean_mfccs = [np.mean(mfcc) for mfcc in mfccs]
print(f'Audio file name: {file_name}')
print(f'Audio file is {len(y)} samples long.')
uid = uuid.uuid4()
doc_ref = db.collection(u'sounds').document(str(uid))
doc_ref.set({
u'uid': str(uid),
u'blob_name': blob_name,
u'file_name': file_name,
u'length': len(y),
u'mean_mfccs': mean_mfccs
})
filters_init = filters.copy()
use_gpu = True
filters_gpu = tf.placeholder(tf.float32, shape=filters.shape, name="Filters")
Rn_gpu = tf.placeholder(tf.float32, shape=(batch_size, filter_size), name="Rn")
prods_gpu = math_ops.matmul(filters_gpu, tf.transpose(Rn_gpu))
gpu_session = tf.Session()
assert not use_gpu or jobs == 1, "Can't use gpu with multiple processes"
data_source = [data_source[1]]
for source_id, source_filename in enumerate(data_source):
data, source_sr = lr.load(source_filename)
data = data[:500000]
data = lr.resample(data, source_sr, target_sr, scale=True)
data_test = lr.resample(data, target_sr, source_sr, scale=True)
lr.output.write_wav("/home/alexeyche/Music/ml/test.wav", data_test, source_sr)
data_denom = np.sqrt(np.sum(data ** 2))
data = data/data_denom
data = np.concatenate([data, np.zeros(filter_size)])
print "Source with id {} and file {}".format(source_id, source_filename)
processes = []
records = []
def sync(wait_all=True):
global dfilters, records, processes, filters
while len(processes)>0:
def __init__(self,
db, # data source
name = '', # optional name
selectors = dict(),
partitioner = None,
meta_sources = [], # optional sources other than 'features' and 'targets' from metadata
channel_filter = NoChannelFilter(), # optional channel filter, default: keep all
channel_names = None, # optional channel names (for metadata)
label_attribute = 'label', # metadata attribute to be used as label
label_map = None, # optional conversion of labels
use_targets = True, # use targets if provides, otherwise labels are used
remove_dc_offset = False, # optional subtraction of channel mean, usually done already earlier
resample = None, # optional down-sampling
normalize = True, # normalize to max=1
# optional sub-sequences selection
start_sample = 0,
stop_sample = None, # optional for selection of sub-sequences
zero_padding = True, # if True (default) trials that are too short will be padded with
# otherwise they will rejected.
# optional signal filter to by applied before splitting the signal
signal_filter = None,
trial_processors = [], # optional processing of the trials
target_processor = None, # optional processing of the targets, e.g. zero-padding
transformers = [], # optional transformations of the dataset
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
debug=False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
self.name = name
self.debug = debug
metadb = DatasetMetaDB(db.metadata, selectors.keys())
if partitioner is not None:
pass # FIXME
selected_trial_ids = metadb.select(selectors)
log.info('selectors: {}'.format(selectors))
log.info('selected trials: {}'.format(selected_trial_ids))
if normalize:
log.info('Data will be normalized to max amplitude 1 per channel (normalize=True).')
trials = list()
labels = list()
targets = list()
meta = list()
if stop_sample == 'auto-min':
stop_sample = np.min([db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using minimum trial length. stop_sample={}'.format(stop_sample))
elif stop_sample == 'auto-max':
stop_sample = np.max([db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using maximum trial length. stop_sample={}'.format(stop_sample))
for trial_i in selected_trial_ids:
trial_meta = db.metadata[trial_i]
if use_targets:
if targets is None:
target = None
else:
target = db.targets[trial_i]
assert not np.isnan(np.sum(target))
if target_processor is not None:
target = target_processor.process(target, trial_meta)
assert not np.isnan(np.sum(target))
else:
# get and process label
label = db.metadata[trial_i][label_attribute]
if label_map is not None:
label = label_map[label]
processed_trial = []
trial = db.data[trial_i]
if np.isnan(np.sum(trial)):
print trial_i, trial
assert not np.isnan(np.sum(trial))
rejected = False # flag for trial rejection
trial = np.atleast_2d(trial)
# process 1 channel at a time
for channel in xrange(trial.shape[0]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = trial[channel, :]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples, resample[0], resample[1], res_type='sinc_best')
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
if stop_sample is not None and stop_sample > len(samples):
if zero_padding:
tmp = np.zeros(stop_sample)
tmp[:len(samples)] = samples
samples = tmp
else:
rejected = True
break # stop processing this trial
s = samples[start_sample:stop_sample]
# TODO optional channel processing
# normalize to max amplitude 1
if normalize:
s = librosa.util.normalize(s)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
s = np.asfarray(s, dtype=theano.config.floatX)
processed_trial.append(s)
### end of channel iteration ###
if rejected:
continue # next trial
processed_trial = np.asfarray([processed_trial], dtype=theano.config.floatX)
# processed_trial = processed_trial.reshape((1, processed_trial.shape))
processed_trial = np.rollaxis(processed_trial, 1, 4)
# optional (external) trial processing, e.g. windowing
# trials will be in b01c format with tf layout for 01-axes
for trial_processor in trial_processors:
processed_trial = trial_processor.process(processed_trial, trial_meta)
trials.append(processed_trial)
for k in range(len(processed_trial)):
meta.append(trial_meta)
if use_targets:
targets.append(target)
else:
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
self.trials = np.vstack(trials)
assert not np.isnan(np.sum(self.trials))
# prepare targets / labels
if use_targets:
self.targets = np.vstack(targets)
assert not np.isnan(np.sum(self.targets))
else:
labels = np.hstack(labels)
if label_map is None:
one_hot_formatter = OneHotFormatter(max(labels) + 1)
else:
one_hot_formatter = OneHotFormatter(max(label_map.values()) + 1)
one_hot_y = one_hot_formatter.format(labels)
self.targets = one_hot_y
self.metadata = meta
if layout == 'ft': # swap axes to (batch, feature, time, channels)
self.trials = self.trials.swapaxes(1, 2)
# transform after finalizing the data structure
for transformer in transformers:
self.trials, self.targets = transformer.process(self.trials, self.targets)
self.trials = np.asarray(self.trials, dtype=theano.config.floatX)
log.debug('final dataset shape: {} (b,0,1,c)'.format(self.trials.shape))
# super(EEGEpochsDataset, self).__init__(topo_view=self.trials, y=self.targets, axes=['b', 0, 1, 'c'])
self.X = self.trials.reshape(self.trials.shape[0], np.prod(self.trials.shape[1:]))
self.y = self.targets
log.info('generated dataset "{}" with shape X={}={} y={} targets={} '.
format(self.name, self.X.shape, self.trials.shape, self.y.shape, self.targets.shape))
# determine data specs
features_space = Conv2DSpace(
shape=[self.trials.shape[1], self.trials.shape[2]],
num_channels=self.trials.shape[3]
)
features_source = 'features'
targets_space = VectorSpace(dim=self.targets.shape[-1])
targets_source = 'targets'
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
# additional support for meta information
self.meta_maps = dict()
for meta_source in meta_sources:
self.meta_maps[meta_source] = sorted(list(set([m[meta_source] for m in self.metadata])))
space_components.extend([VectorSpace(dim=1)])
source_components.extend([meta_source])
log.info('Generated meta-source "{}" with value map: {}'
.format(meta_source, self.meta_maps[meta_source]))
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
log.debug('data specs: {}'.format(self.data_specs))
# tvars = tf.trainable_variables()
# grads_raw = tf.gradients(cost, tvars)
# grads, _ = tf.clip_by_global_norm(grads_raw, 5.0)
# apply_grads = optimizer.apply_gradients(zip(grads, tvars))
##################################
# DATA
fname = env.dataset([f for f in os.listdir(env.dataset()) if f.endswith(".wav")][0])
df = env.run("test_data.pkl")
if not os.path.exists(df):
song_data_raw, source_sr = lr.load(fname)
print "Got sampling rate {}, resampling to {} ...".format(source_sr, target_sr)
song_data = lr.resample(song_data_raw, source_sr, target_sr, scale=True)
song_data = song_data[:30000,]
np.save(open(df, "w"), song_data)
else:
song_data = np.load(open(df))
inputs_v, data_denom = norm(song_data)
##################################
# EVALUATION
sess = tf.Session()
def __init__(self,
path,
name = '', # optional name
# selectors
subjects='all', # optional selector (list) or 'all'
trial_types='all', # optional selector (list) or 'all'
trial_numbers='all', # optional selector (list) or 'all'
conditions='all', # optional selector (list) or 'all'
partitioner = None,
channel_filter = NoChannelFilter(), # optional channel filter, default: keep all
channel_names = None, # optional channel names (for metadata)
label_map = None, # optional conversion of labels
remove_dc_offset = False, # optional subtraction of channel mean, usually done already earlier
resample = None, # optional down-sampling
# optional sub-sequences selection
start_sample = 0,
stop_sample = None, # optional for selection of sub-sequences
# optional signal filter to by applied before spitting the signal
signal_filter = None,
# windowing parameters
frame_size = -1,
hop_size = -1, # values > 0 will lead to windowing
hop_fraction = None, # alternative to specifying absolute hop_size
# optional spectrum parameters, n_fft = 0 keeps raw data
n_fft = 0,
n_freq_bins = None,
spectrum_log_amplitude = False,
spectrum_normalization_mode = None,
include_phase = False,
flatten_channels=False,
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
save_matrix_path = None,
keep_metadata = False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
# TODO: get the whole filtering into an extra class
datafiles_metadata, metadb = load_datafiles_metadata(path)
# print datafiles_metadata
def apply_filters(filters, node):
if isinstance(node, dict):
filtered = []
keepkeys = filters[0]
for key, value in node.items():
if keepkeys == 'all' or key in keepkeys:
filtered.extend(apply_filters(filters[1:], value))
return filtered
else:
return node # [node]
# keep only files that match the metadata filters
self.datafiles = apply_filters([subjects,trial_types,trial_numbers,conditions], datafiles_metadata)
# copy metadata for retained files
self.metadb = {}
for datafile in self.datafiles:
self.metadb[datafile] = metadb[datafile]
# print self.datafiles
# print self.metadb
self.name = name
if partitioner is not None:
self.datafiles = partitioner.get_partition(self.name, self.metadb)
self.include_phase = include_phase
self.spectrum_normalization_mode = spectrum_normalization_mode
self.spectrum_log_amplitude = spectrum_log_amplitude
self.sequence_partitions = [] # used to keep track of original sequences
# metadata: [subject, trial_no, stimulus, channel, start, ]
self.metadata = []
sequences = []
labels = []
n_sequences = 0
if frame_size > 0 and hop_size == -1 and hop_fraction is not None:
hop_size = np.ceil(frame_size / hop_fraction)
for i in xrange(len(self.datafiles)):
with log_timing(log, 'loading data from {}'.format(self.datafiles[i])):
# save start of next sequence
self.sequence_partitions.append(n_sequences)
data, metadata = load(os.path.join(path, self.datafiles[i]))
label = metadata['label']
if label_map is not None:
label = label_map[label]
multi_channel_frames = []
# process 1 channel at a time
for channel in xrange(data.shape[1]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = data[:, channel]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples, resample[0], resample[1])
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
samples = samples[start_sample:stop_sample]
if n_fft is not None and n_fft > 0: # Optionally:
### frequency spectrum branch ###
# transform to spectogram
hop_length = n_fft / 4;
'''
from http://theremin.ucsd.edu/~bmcfee/librosadoc/librosa.html
>>> # Get a power spectrogram from a waveform y
>>> S = np.abs(librosa.stft(y)) ** 2
>>> log_S = librosa.logamplitude(S)
'''
S = librosa.core.stft(samples, n_fft=n_fft, hop_length=hop_length)
# mag = np.abs(S) # magnitude spectrum
mag = np.abs(S)**2 # power spectrum
# include phase information if requested
if self.include_phase:
# phase = np.unwrap(np.angle(S))
phase = np.angle(S)
# Optionally: cut off high bands
if n_freq_bins is not None:
mag = mag[0:n_freq_bins, :]
if self.include_phase:
phase = phase[0:n_freq_bins, :]
if self.spectrum_log_amplitude:
mag = librosa.logamplitude(mag)
s = mag # for normalization
'''
NOTE on normalization:
It depends on the structure of a neural network and (even more)
on the properties of data. There is no best normalization algorithm
because if there would be one, it would be used everywhere by default...
In theory, there is no requirement for the data to be normalized at all.
This is a purely practical thing because in practice convergence could
take forever if your input is spread out too much. The simplest would be
to just normalize it by scaling your data to (-1,1) (or (0,1) depending
on activation function), and in most cases it does work. If your
algorithm converges well, then this is your answer. If not, there are
too many possible problems and methods to outline here without knowing
the actual data.
'''
## normalize to mean 0, std 1
if self.spectrum_normalization_mode == 'mean0_std1':
# s = preprocessing.scale(s, axis=0);
mean = np.mean(s)
std = np.std(s)
s = (s - mean) / std
## normalize by linear transform to [0,1]
elif self.spectrum_normalization_mode == 'linear_0_1':
s = s / np.max(s)
## normalize by linear transform to [-1,1]
elif self.spectrum_normalization_mode == 'linear_-1_1':
s = -1 + 2 * (s - np.min(s)) / (np.max(s) - np.min(s))
elif self.spectrum_normalization_mode is not None:
raise ValueError(
'unsupported spectrum normalization mode {}'.format(
self.spectrum_normalization_mode)
)
#print s.mean(axis=0)
#print s.std(axis=0)
# include phase information if requested
if self.include_phase:
# normalize phase to [-1.1]
phase = phase / np.pi
s = np.vstack([s, phase])
# transpose to fit pylearn2 layout
s = np.transpose(s)
# print s.shape
### end of frequency spectrum branch ###
else:
### raw waveform branch ###
# normalize to max amplitude 1
s = librosa.util.normalize(samples)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
### end of raw waveform branch ###
s = np.asfarray(s, dtype='float32')
if frame_size > 0 and hop_size > 0:
s = s.copy() # FIXME: THIS IS NECESSARY IN MultiChannelEEGSequencesDataset - OTHERWISE, THE FOLLOWING OP DOES NOT WORK!!!!
frames = frame(s, frame_length=frame_size, hop_length=hop_size)
else:
frames = s
del s
# print frames.shape
if flatten_channels:
# add artificial channel dimension
frames = frames.reshape((frames.shape[0], frames.shape[1], frames.shape[2], 1))
# print frames.shape
sequences.append(frames)
# increment counter by new number of frames
n_sequences += frames.shape[0]
if keep_metadata:
# determine channel name
channel_name = None
if channel_names is not None:
channel_name = channel_names[channel]
elif 'channels' in metadata:
channel_name = metadata['channels'][channel]
self.metadata.append({
'subject' : metadata['subject'], # subject
'trial_type': metadata['trial_type'], # trial_type
'trial_no' : metadata['trial_no'], # trial_no
'condition' : metadata['condition'], # condition
'channel' : channel, # channel
'channel_name' : channel_name,
'start' : self.sequence_partitions[-1], # start
'stop' : n_sequences # stop
})
for _ in xrange(frames.shape[0]):
labels.append(label)
else:
multi_channel_frames.append(frames)
### end of channel iteration ###
if not flatten_channels:
# turn list into array
multi_channel_frames = np.asfarray(multi_channel_frames, dtype='float32')
# [channels x frames x time x freq] -> cb01
# [channels x frames x time x 1] -> cb0.
# move channel dimension to end
multi_channel_frames = np.rollaxis(multi_channel_frames, 0, 4)
# print multi_channel_frames.shape
# log.debug(multi_channel_frames.shape)
sequences.append(multi_channel_frames)
# increment counter by new number of frames
n_sequences += multi_channel_frames.shape[0]
if keep_metadata:
self.metadata.append({
'subject' : metadata['subject'], # subject
'trial_type': metadata['trial_type'], # trial_type
'trial_no' : metadata['trial_no'], # trial_no
'condition' : metadata['condition'], # condition
'channel' : 'all', # channel
'start' : self.sequence_partitions[-1], # start
'stop' : n_sequences # stop
})
for _ in xrange(multi_channel_frames.shape[0]):
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
sequences = np.vstack(sequences)
# print sequences.shape;
labels = np.hstack(labels)
# one_hot_y = one_hot(labels)
one_hot_formatter = OneHotFormatter(labels.max() + 1) # FIXME!
one_hot_y = one_hot_formatter.format(labels)
self.labels = labels
if layout == 'ft': # swap axes to (batch, feature, time, channels)
sequences = sequences.swapaxes(1, 2)
log.debug('final dataset shape: {} (b,0,1,c)'.format(sequences.shape))
super(MultiChannelEEGDataset, self).__init__(topo_view=sequences, y=one_hot_y, axes=['b', 0, 1, 'c'])
log.info('generated dataset "{}" with shape X={}={} y={} labels={} '.
format(self.name, self.X.shape, sequences.shape, self.y.shape, self.labels.shape))
if save_matrix_path is not None:
matrix = DenseDesignMatrix(topo_view=sequences, y=one_hot_y, axes=['b', 0, 1, 'c'])
with log_timing(log, 'saving DenseDesignMatrix to {}'.format(save_matrix_path)):
serial.save(save_matrix_path, matrix)