def load_spectrograms(filename):
train_data = []
spectrogram, sp_time, sp_freq, fs = ap.get_spectrogram(
filename, ST_WIN, ST_STEP)
# These should change depending on the signal's size
spec_resize_ratio_freq = 4
spec_resize_ratio_time = 4
f_low = F1 if F1 < fs / 2.0 else fs / 2.0
f_high = F2 if F2 < fs / 2.0 else fs / 2.0
# define feature sequence for vocalization detection
f1 = np.argmin(np.abs(sp_freq - f_low))
f2 = np.argmin(np.abs(sp_freq - f_high))
spectral_energy_1 = spectrogram.sum(axis=1)
spectral_energy_2 = spectrogram[:, f1:f2].sum(axis=1)
seg_limits, thres_sm, _ = ap.get_syllables(spectral_energy_2,
spectral_energy_1,
ST_STEP,
threshold_per=thres * 100,
min_duration=MIN_VOC_DUR)
train_data += (ar.cluster_syllables(seg_limits,
spectrogram,
sp_freq,
f_low,
f_high,
ST_STEP,
train=True))
return train_data
def signal_handler(signal, frame):
"""
This function is called when Ctr + C is pressed and is used to output the
final buffer into a WAV file
"""
# write final buffer to wav file
global fs
if len(all_data) > 1:
wavfile.write(outstr + ".wav", fs, np.int16(all_data))
spectrogram, sp_time, sp_freq, fs = ap.get_spectrogram(
outstr + ".wav", ST_WIN, ST_STEP)
f_low = F1 if F1 < fs / 2.0 else fs / 2.0
f_high = F2 if F2 < fs / 2.0 else fs / 2.0
# define feature sequence for vocalization detection
f1 = np.argmin(np.abs(sp_freq - f_low))
f2 = np.argmin(np.abs(sp_freq - f_high))
spectral_energy, means, max_values = ap.prepare_features(
spectrogram[:, f1:f2])
time_sec = 100
seg_limits, thres_sm = ap.get_syllables(
spectral_energy,
means,
max_values,
ST_STEP,
threshold_per=thres * 100,
min_duration=MIN_VOC_DUR,
threshold_buf=means,
)
for s in seg_limits:
with open("debug_offline.csv", "a") as fp:
fp.write(f'{count_mid_bufs * mid_buffer_size + s[0]},'
f'{count_mid_bufs * mid_buffer_size + s[1]}\n')
sys.exit(0)
"--ground_truth_file",
required=True,
nargs=None,
help="Ground truth file")
return parser.parse_args()
if __name__ == "__main__":
args = parse_arguments()
thres = args.threshold
spec_resize_ratio_freq = args.resize_freq
spec_resize_ratio_time = args.resize_time
# feature (spectrogram) extraction:
spectrogram, sp_time, sp_freq, fs = ap.get_spectrogram(
args.input_file, args.win, args.step)
duration = spectrogram.shape[0] * args.step
f_low = F1 if F1 < fs / 2.0 else fs / 2.0
f_high = F2 if F2 < fs / 2.0 else fs / 2.0
# define feature sequence for vocalization detection
f1 = np.argmin(np.abs(sp_freq - f_low))
f2 = np.argmin(np.abs(sp_freq - f_high))
spectral_energy_1 = spectrogram.sum(axis=1)
spectral_energy_2 = spectrogram[:, f1:f2].sum(axis=1)
segs, thres_sm, spectral_ratio = ap.get_syllables(spectral_energy_2,
spectral_energy_1,
args.step,
dbc.Row(id='intermediate_val_total_clusters',
style={'display': 'none'}),
dbc.Row(id='intermediate_val_clusters',
style={'display': 'none'}),
dbc.Row(id='clustering_info', style={'display': 'none'})
],
style={"height": "100vh"})
return layout
if __name__ == "__main__":
args = parse_arguments()
global sp_time, sp_freq, moves, click_index
click_index = -1
time_start = time.time()
spectrogram, sp_time, sp_freq, fs = ap.get_spectrogram(
args.input_file, ST_WIN, ST_STEP)
with open(
'annotations_eval_{}.json'.format(
(args.input_file.split('/')[-1]).split('.')[0]),
'w') as outfile:
x = json.dumps(
{
'input_file': args.input_file.split('/')[-1],
'total_cluster_annotations': [],
'cluster_annotations': [],
'point_annotations': []
},
indent=4)
outfile.write(x)
# These should change depending on the signal's size