def plot_data(audio_files,frame, window_type, nfft, step, fmin, fmax, chunk = None, detections=None, detections_channel=0):
graph_spectros = GrapherFactory('SoundPlotter', title='Spectrograms', frequency_max=fmax)
graph_waveforms = GrapherFactory('SoundPlotter', title='Waveforms')
for audio_file, channel in zip(audio_files['path'], audio_files['channel'] ): # for each channel
# load waveform
sound = Sound(audio_file)
sound.read(channel=channel, chunk=chunk, unit='sec', detrend=True)
# Calculates spectrogram
spectro = Spectrogram(frame, window_type, nfft, step, sound.waveform_sampling_frequency, unit='sec')
spectro.compute(sound, dB=True, use_dask=False)
# Crop unused frequencies
spectro.crop(frequency_min=fmin, frequency_max=fmax, inplace=True)
# Plot
graph_spectros.add_data(spectro, time_offset_sec=chunk[0])
graph_waveforms.add_data(sound, time_offset_sec=chunk[0])
graph_spectros.colormap = 'binary'
if detections:
graph_spectros.add_annotation(detections, panel=detections_channel, color='green',label='Detections')
graph_waveforms.add_annotation(detections, panel=detections_channel, color='green',label='Detections')
if chunk:
graph_spectros.time_min = chunk[0]
graph_spectros.time_max = chunk[1]
graph_waveforms.time_min = chunk[0]
graph_waveforms.time_max = chunk[1]
graph_spectros.show()
graph_waveforms.show()
def plot_spectrogram(audio_file,loc,t1_sec, t2_sec, geometry=(1,1,1)):
fmin=0
fmax=1000
frame= 0.0625
window_type= 'hann'
nfft=0.0853
step=0.01
channel=0
chunk=[t1_sec,t2_sec]
graph_spectros = GrapherFactory('SoundPlotter', title='Spectrograms', frequency_max=fmax)
sound = Sound(audio_file)
sound.read(channel=channel, chunk=chunk, unit='sec', detrend=True)
# Calculates spectrogram
spectro = Spectrogram(frame, window_type, nfft, step, sound.waveform_sampling_frequency, unit='sec')
spectro.compute(sound, dB=True, use_dask=False)
# Crop unused frequencies
spectro.crop(frequency_min=fmin, frequency_max=fmax, inplace=True)
# Plot
graph_spectros.add_data(spectro)
#graph_spectros.add_annotation(loc, panel=0, color='burlywood',label='Detections')
graph_spectros.add_annotation(loc, panel=0, color='peachpuff')
graph_spectros.colormap = 'binary' #'jet'
fig, ax = graph_spectros.show()
if ax.get_geometry() != geometry :
ax.change_geometry(*geometry)
return fig, ax
def stack_waveforms(audio_files, detec, TDOA_max_sec):
waveform_stack = []
for audio_file, channel in zip(audio_files['path'], audio_files['channel'] ): # for each channel
# load waveform
chan_wav = Sound(audio_file)
chan_wav.read(channel=channel,
chunk=[detec['time_min_offset']-TDOA_max_sec, detec['time_max_offset']+TDOA_max_sec],
unit='sec',
detrend=True)
# bandpass filter
chan_wav.filter('bandpass', [detec['frequency_min'], detec['frequency_max']])
# stack
waveform_stack.append(chan_wav.waveform)
return waveform_stack
def run_detector(infile, outdir, classif_model=None, deployment_file=None):
## Input paraneters ##########################################################
# Spectrogram parameters
frame = 0.0625 #3000
nfft = 0.0853 # 4096
step = 0.01 # 5
fmin = 0
fmax = 1000
window_type = 'hann'
# start and stop time of wavfile to analyze
#t1 = 0 # 24
#t2 = 60 # 40
## ###########################################################################
outfile = os.path.join(outdir, os.path.split(infile)[1] + '.nc')
if os.path.exists(outfile) is False:
# load audio data
sound = Sound(infile)
#sound.read(channel=0, chunk=[t1, t2], unit='sec')
sound.read(channel=0, unit='sec')
# Calculates spectrogram
print('Spectrogram')
spectro = Spectrogram(frame,
window_type,
nfft,
step,
sound.waveform_sampling_frequency,
unit='sec')
spectro.compute(sound, dB=True, use_dask=True, dask_chunks=100)
# Crop unused frequencies
spectro.crop(frequency_min=fmin, frequency_max=fmax, inplace=True)
# Denoise
print('Denoise')
spectro.denoise(
'median_equalizer',
window_duration=3,
use_dask=True,
dask_chunks=(50, 50000), #'auto',#(87,10000),
inplace=True)
# Detector
print('Detector')
file_timestamp = ecosound.core.tools.filename_to_datetime(infile)[0]
detector = DetectorFactory('BlobDetector',
kernel_duration=0.1,
kernel_bandwidth=300,
threshold=10,
duration_min=0.05,
bandwidth_min=40)
detections = detector.run(
spectro,
start_time=file_timestamp,
use_dask=True,
dask_chunks=(4096, 50000), #'auto',
debug=False)
# Maasurements
print('Measurements')
spectro_features = MeasurerFactory('SpectrogramFeatures',
resolution_time=0.001,
resolution_freq=0.1,
interp='linear')
measurements = spectro_features.compute(spectro,
detections,
debug=False,
verbose=False,
use_dask=True)
# Add metadata
if deployment_file:
measurements.insert_metadata(deployment_file)
# Add file informations
file_name = os.path.splitext(os.path.basename(infile))[0]
file_dir = os.path.dirname(infile)
file_ext = os.path.splitext(infile)[1]
measurements.insert_values(
operator_name=platform.uname().node,
audio_file_name=file_name,
audio_file_dir=file_dir,
audio_file_extension=file_ext,
audio_file_start_date=ecosound.core.tools.filename_to_datetime(
infile)[0])
# Classification
print('Classification')
if classif_model:
features = classif_model['features']
model = classif_model['model']
Norm_mean = classif_model['normalization_mean']
Norm_std = classif_model['normalization_std']
classes_encoder = classif_model['classes']
# data dataframe
data = measurements.data
n1 = len(data)
# drop observations/rows with NaNs
data = data.replace([np.inf, -np.inf], np.nan)
data.dropna(subset=features,
axis=0,
how='any',
thresh=None,
inplace=True)
n2 = len(data)
print('Deleted observations (due to NaNs): ' + str(n1 - n2))
# Classification - predictions
X = data[features]
X = (X - Norm_mean) / Norm_std
pred_class = model.predict(X)
pred_prob = model.predict_proba(X)
pred_prob = pred_prob[range(0, len(pred_class)), pred_class]
# Relabel
for index, row in classes_encoder.iterrows():
pred_class = [
row['label'] if i == row['ID'] else i for i in pred_class
]
# update measurements
data['label_class'] = pred_class
data['confidence'] = pred_prob
# sort detections by ascending start date/time
data.sort_values('time_min_offset',
axis=0,
ascending=True,
inplace=True)
# save result as NetCDF file
print('Saving')
measurements.data = data
measurements.to_netcdf(outfile)
else:
print('Recording already processed.')
frame = 3000
nfft = 4096
step = 500
#ovlp = 2500
fmin = 0
fmax = 1000
window_type = 'hann'
# start and stop time of wavfile to analyze
t1 = 1
t2 = 50
## ###########################################################################
tic = time.perf_counter()
# load audio data
sound = Sound(single_channel_file)
sound.read(channel=0, chunk=[t1, t2], unit='sec', detrend=True)
# Calculates spectrogram
spectro = Spectrogram(frame,
window_type,
nfft,
step,
sound.waveform_sampling_frequency,
unit='samp')
spectro.compute(sound, dB=True, use_dask=True, dask_chunks=40)
# Crop unused frequencies
spectro.crop(frequency_min=fmin, frequency_max=fmax, inplace=True)
# Denoise
def run_detector(infile, channel, config, chunk=None, deployment_file=None):
sound = Sound(infile)
# load audio data
if chunk:
sound.read(channel=channel, chunk=[t1, t2], unit='sec', detrend=True)
time_offset_sec = t1
else:
sound.read(channel=channel, detrend=True)
time_offset_sec = 0
# Calculates spectrogram
spectro = Spectrogram(config['SPECTROGRAM']['frame_sec'],
config['SPECTROGRAM']['window_type'],
config['SPECTROGRAM']['nfft_sec'],
config['SPECTROGRAM']['step_sec'],
sound.waveform_sampling_frequency,
unit='sec',)
spectro.compute(sound,
config['SPECTROGRAM']['dB'],
config['SPECTROGRAM']['use_dask'],
config['SPECTROGRAM']['dask_chunks'],)
spectro.crop(frequency_min=config['SPECTROGRAM']['fmin_hz'],
frequency_max=config['SPECTROGRAM']['fmax_hz'],
inplace=True,
)
# Denoise
spectro.denoise(config['DENOISER']['denoiser_name'],
window_duration=config['DENOISER']['window_duration_sec'],
use_dask=config['DENOISER']['use_dask'],
dask_chunks=tuple(config['DENOISER']['dask_chunks']),
inplace=True)
# Detector
file_timestamp = ecosound.core.tools.filename_to_datetime(sound.file_full_path)[0]
detector = DetectorFactory(config['DETECTOR']['detector_name'],
kernel_duration=config['DETECTOR']['kernel_duration_sec'],
kernel_bandwidth=config['DETECTOR']['kernel_bandwidth_hz'],
threshold=config['DETECTOR']['threshold'],
duration_min=config['DETECTOR']['duration_min_sec'],
bandwidth_min=config['DETECTOR']['bandwidth_min_hz']
)
start_time = file_timestamp + datetime.timedelta(seconds=time_offset_sec)
detections = detector.run(spectro,
start_time=start_time,
use_dask=config['DETECTOR']['use_dask'],
dask_chunks=tuple(config['DETECTOR']['dask_chunks']),
debug=False,
)
# add time offset in only a section of recording was analysed.
detections.data['time_min_offset'] = detections.data['time_min_offset'] + time_offset_sec
detections.data['time_max_offset'] = detections.data['time_max_offset'] + time_offset_sec
# add deployment metadata
detections.insert_metadata(deployment_file, channel=channel)
# Add file informations
file_name = os.path.splitext(os.path.basename(sound.file_full_path))[0]
file_dir = os.path.dirname(sound.file_full_path)
file_ext = os.path.splitext(sound.file_full_path)[1]
detections.insert_values(operator_name=platform.uname().node,
audio_file_name=file_name,
audio_file_dir=file_dir,
audio_file_extension=file_ext,
audio_file_start_date=ecosound.core.tools.filename_to_datetime(sound.file_full_path)[0]
)
return detections
def run_detector(infile, outdir, classif_model, config, deployment_file=None, extension=".wav", overwrite=False, netcdf= True, pamlab=False, raven=False):
"""
Run the fish sound detector.
Parameters
----------
infile : str
Path of the audio file to process.
outdir : str
Path of the output folder where the results will be written.
classif_model : str
Path and name of the classification model to use (.sav pickle file)
config : dict
Dict with all parameters from the yaml file.
deployment_file : str, optional
Path and name of the csv file with all the deployment information.
The default is None.
extension : str, optional
Extension of the audio files to process. The default is ".wav".
overwrite : bool, optional
If set to True, overwrites results (i.e. netcdf files) even if they
already exist in the outdir folder. The default is False.
netcdf : bool, optional
If set to True, saves results as netcdf4 files (.nc).
The default is True.
pamlab : bool, optional
If set to True, saves results as PAMlab files (.log).
The default is False.
raven : bool, optional
If set to True, saves results as Raven files (.txt).
The default is False.
Returns
-------
None.
"""
outfile = os.path.join(outdir, os.path.split(infile)[1] + '.nc')
if (os.path.exists(outfile) is False) or (os.path.exists(outfile) and overwrite):
# load audio data
sound = Sound(infile)
sound.read(channel=config['AUDIO']['channel'], unit='sec')
# Calculates spectrogram
print('Spectrogram')
spectro = Spectrogram(config['SPECTROGRAM']['frame_sec'],
config['SPECTROGRAM']['window_type'],
config['SPECTROGRAM']['nfft_sec'],
config['SPECTROGRAM']['step_sec'],
sound.waveform_sampling_frequency,
unit='sec',
)
spectro.compute(sound,
config['SPECTROGRAM']['dB'],
config['SPECTROGRAM']['use_dask'],
config['SPECTROGRAM']['dask_chunks'],
)
# Crop unused frequencies
spectro.crop(frequency_min=config['SPECTROGRAM']['fmin_hz'],
frequency_max=config['SPECTROGRAM']['fmax_hz'],
inplace=True,
)
# Denoise
print('Denoise')
spectro.denoise(config['DENOISER']['denoiser_name'],
window_duration=config['DENOISER']['window_duration_sec'],
use_dask=config['DENOISER']['use_dask'],
dask_chunks=tuple(config['DENOISER']['dask_chunks']),
inplace=True)
# Detector
print('Detector')
file_timestamp = ecosound.core.tools.filename_to_datetime(infile)[0]
detector = DetectorFactory(config['DETECTOR']['detector_name'],
kernel_duration=config['DETECTOR']['kernel_duration_sec'],
kernel_bandwidth=config['DETECTOR']['kernel_bandwidth_hz'],
threshold=config['DETECTOR']['threshold'],
duration_min=config['DETECTOR']['duration_min_sec'],
bandwidth_min=config['DETECTOR']['bandwidth_min_hz']
)
detections = detector.run(spectro,
start_time=file_timestamp,
use_dask=config['DETECTOR']['use_dask'],
dask_chunks=tuple(config['DETECTOR']['dask_chunks']),
debug=False,
)
# Maasurements
print('Measurements')
spectro_features = MeasurerFactory(config['MEASURER']['measurer_name'],
resolution_time=config['MEASURER']['resolution_time_sec'],
resolution_freq=config['MEASURER']['resolution_freq_hz'],
interp=config['MEASURER']['interp'],
)
measurements = spectro_features.compute(spectro,
detections,
debug=False,
verbose=False,
use_dask=config['MEASURER']['use_dask'])
# Add metadata
if deployment_file:
measurements.insert_metadata(deployment_file)
else:
measurements.insert_values(audio_channel=0,
UTC_offset=0,
audio_sampling_frequency=0,
audio_bit_depth=0,
mooring_platform_name='',
recorder_type='',
recorder_SN='',
hydrophone_model='',
hydrophone_SN='',
hydrophone_depth=0,
location_name='',
location_lat=0,
location_lon=0,
location_water_depth=0,
deployment_ID='',
)
# Add file informations
file_name = os.path.splitext(os.path.basename(infile))[0]
file_dir = os.path.dirname(infile)
file_ext = os.path.splitext(infile)[1]
measurements.insert_values(operator_name=platform.uname().node,
audio_file_name=file_name,
audio_file_dir=file_dir,
audio_file_extension=file_ext,
audio_file_start_date=ecosound.core.tools.filename_to_datetime(infile)[0]
)
# Classification
print('Classification')
if classif_model:
features = classif_model['features']
model = classif_model['model']
Norm_mean = classif_model['normalization_mean']
Norm_std = classif_model['normalization_std']
classes_encoder = classif_model['classes']
# data dataframe
data = measurements.data
n1 = len(data)
# drop observations/rows with NaNs
data = data.replace([np.inf, -np.inf], np.nan)
data.dropna(subset=features,
axis=0,
how='any',
thresh=None,
inplace=True)
n2 = len(data)
print('Deleted observations (due to NaNs): ' + str(n1-n2))
# Classification - predictions
X = data[features]
X = (X-Norm_mean)/Norm_std
pred_class = model.predict(X)
pred_prob = model.predict_proba(X)
pred_prob = pred_prob[range(0, len(pred_class)), pred_class]
# Relabel
for index, row in classes_encoder.iterrows():
pred_class = [row['label'] if i == row['ID'] else i for i in pred_class]
# update measurements
data['label_class'] = pred_class
data['confidence'] = pred_prob
# sort detections by ascending start date/time
data.sort_values('time_min_offset',
axis=0,
ascending=True,
inplace=True)
# save result as NetCDF file
print('Saving')
measurements.data = data
if netcdf:
measurements.to_netcdf(outfile)
if pamlab:
measurements.to_pamlab(outdir)
if raven:
measurements.to_raven(outdir)
else:
print('Recording already processed.')
logging.info('Recording already processed.')
frame = 3000
nfft = 4096
step = 500
#ovlp = 2500
fmin = 0
fmax = 1000
window_type = 'hann'
# start and stop time of wavfile to analyze
t1 = 1515
t2 = 1541
## ###########################################################################
tic = time.perf_counter()
# load audio data
sound = Sound(single_channel_file)
fs = sound.waveform_sampling_frequency
sound.read(channel=0, chunk=[round(t1 * fs), round(t2 * fs)])
# Calculates spectrogram
Spectro = Spectrogram(frame, window_type, nfft, step, fs, unit='samp')
Spectro.compute(sound, dB=True, dask=True)
toc = time.perf_counter()
print(f"Executed in {toc - tic:0.4f} seconds")
# # Plot
# graph = GrapherFactory('SoundPlotter', title='Recording', frequency_max=1000)
# graph.add_data(sound)
# graph.add_data(Spectro)
# graph.colormap = 'jet'
print(f"Dexecuted in {toc - tic:0.4f} seconds")
return axis_frequencies, axis_times, spectrogram
if __name__ == '__main__':
# # Create random signal
# fs = 48000
# sig_dur = 60#1800
# sig = np.random.rand(sig_dur*fs)
single_channel_file = r"../ecosound/resources/67674121.181018013806.wav"
t1 = 24
t2 = 120
sound = Sound(single_channel_file)
#sound.read(channel=0, chunk=[t1, t2], unit='sec')
sound.read(channel=0)
fs = sound.file_sampling_frequency
sig = sound.waveform
sig = sig - np.mean(sig)
# Calculates spectrogram
frame_samp = 3000
overlap_samp = 2500
fft_samp = 4096
#S1 = spectro_numpy(sig, fs, frame_samp, overlap_samp, fft_samp)
#S2 = spectro_loop(sig, fs, frame_samp, overlap_samp, fft_samp)
#S3 = spectro_loop_dask(sig, fs, frame_samp, overlap_samp, fft_samp)
F, T, S3 = spectro_loop_dask2(
frame = 3000 nfft = 4096 step = 500 #ovlp = 2500 fmin = 0 fmax = 1000 window_type = 'hann' # start and stop time of wavfile to analyze t1 = 0#24 t2 = 200#40 ## ########################################################################### tic = time.perf_counter() # load audio data sound = Sound(audio_file) sound.read(channel=0, chunk=[t1, t2], unit='sec', detrend=True) # Calculates spectrogram spectro = Spectrogram(frame, window_type, nfft, step, sound.waveform_sampling_frequency, unit='samp') spectro.compute(sound, dB=True, use_dask=True, dask_chunks=40) spectro.crop(frequency_min=fmin, frequency_max=fmax, inplace=True) # load annotations annot = Annotation() annot.from_raven(annotation_file) # load detections detec = Measurement() detec.from_netcdf(detection_file)
Created on Thu Feb 6 22:35:31 2020
@author: xavier.mouy
"""
import sys
sys.path.append("..") # Adds higher directory to python modules path.
from ecosound.core.audiotools import Sound, Filter
#single_channel_file = r"../ecosound/resources/AMAR173.4.20190916T061248Z.wav"
multi_channel_file = r"../ecosound/resources/671404070.190722162836.wav"
# load part of the file and plot
print('------------------------------')
#sig = Sound() # should return error
sig = Sound(multi_channel_file)
print(len(sig))
sig.read(channel=0, chunk=[10, 1000])
print('------------------------------')
print(len(sig))
print('start sample: ', sig.waveform_start_sample)
print('stop sample: ', sig.waveform_stop_sample)
print('duration:: ', sig.waveform_duration_sample)
print(len(sig))
sig.plot_waveform(newfig=True)
# extract a sinppet from the data
sig2 = sig.select_snippet([100,1000])
sig2.plot_waveform(newfig=True)
print('------------------------------')
print('start sample: ', sig2.waveform_start_sample)