rois = read_audacity_annot(
'./data/spinetail.txt') ## annotations using Audacity
###=============== compute spectrogram =================
Sxx, tn, fn, ext = spectrogram(s, fs)
Sxx = 10 * np.log10(Sxx)
rois = format_features(rois, tn, fn)
###=============== from Audacity =================
### with all labels
ax, fig = overlay_rois(Sxx, ext, rois, vmin=-120, vmax=20)
# Compute an visualize features
shape, params = shape_features(Sxx, resolution='low', rois=rois)
plot_shape(shape.mean(), params)
# Compute and visualize centroids
centroid = centroid_features(Sxx, rois)
centroid = format_features(centroid, tn, fn)
ax, fig = overlay_centroid(Sxx,
ext,
centroid,
savefig=None,
vmin=-120,
vmax=20,
fig=fig,
ax=ax)
###=============== find ROI 2D =================
# Format ROIs and visualize the bounding box on the audio spectrogram.
df_rois = format_features(df_rois, tn, fn)
ax0, fig0 = overlay_rois(Sxx_db, df_rois, **{
'vmin': 0,
'vmax': 60,
'extent': ext
})
#%%
# 2. Compute acoustic features
# ----------------------------
# The ``shape_feaures`` function uses bidimensional wavelets to get the texture and spectro-temporal shape coeficients of each ROI. Wavelets have the advantage of being robust when the signal-to-noise ratio is low, and derive homogeneous descriptors which facilitate the clustering process. The wavelet decomposition is performed on the complete spectrogram, hence the coeficients for ROIs do not vary much even when not the time-frequency bounds are not exact. The centroid features gives an estimate of the median frequency of the ROIs.
df_shape, params = features.shape_features(Sxx_db,
resolution='low',
rois=df_rois)
df_centroid = features.centroid_features(Sxx_db, df_rois)
# Get median frequency and normalize
median_freq = fn[np.round(df_centroid.centroid_y).astype(int)]
df_centroid['centroid_freq'] = median_freq / fn[-1]
#%%
# 3. Reduce the dimensionality of the features
# --------------------------------------------
# The shape audio features have 26 dimensions. To facilitate the clustering process and visualize the results, it is posible to use non-metric dimensionality reduction algorithm, namely the t-distributed stochastic neighbor embedding (t-SNE), to proyect the data in two dimensions.
from sklearn.manifold import TSNE
X = df_shape.loc[:, df_shape.columns.str.startswith('shp')]
def batch_feature_rois_no_verb(rois_list, params_features, path_audio):
"""
Computes features for a list of files
Parameters:
----------
params_features: dict
Dictionary with the basic parameters to feed find_rois:
'flims', 'tlen', and 'th'.
path_flist : str
Path to a *.txt file with the list of audio filenames to process
path_audio : str
Path to the place were the dataset of audio files are stored
path_save : str
Path with the file name to save the csv
Returns:
-------
info_features: dic
Dictionary with features and all the parameters used to compute the features.
Included keys: features, parameters_df, opt_shape, opt_spectro
"""
## TODO: when the time limits are too short, the function has problems
# load parameters
flims = params_features['flims']
opt_spec = params_features['opt_spec']
opt_shape = opt_shape_presets(params_features['opt_shape_str'])
# load detection data
features = []
for idx, file in enumerate(rois_list):
# unpack file values
fname = file['fname']
rois_tf = file['rois']
#print(idx+1, '/', len(rois_list), fname)
if rois_tf.empty:
#print('< No detection on file >')
features.append({'fname':fname, 'features': pd.DataFrame()})
else:
# load materials: sound, spectrogram
s, fs = sound.load(path_audio+fname)
im, dt, df, ext = sound.spectrogram(s, fs, nperseg=opt_spec['nperseg'],
overlap=opt_spec['overlap'], fcrop=flims,
rescale=False, db_range=opt_spec['db_range'])
# format rois to bbox
ts = np.arange(ext[0], ext[1], dt)
f = np.arange(ext[2],ext[3]+df,df)
rois_bbox = format_rois(rois_tf, ts, f, fmt='bbox')
# roi to image blob
im_blobs = rois_to_imblobs(np.zeros(im.shape), rois_bbox)
# get features: shape, center frequency
im = normalize_2d(im, 0, 1)
bbox, params, shape = shape_features(im, im_blobs, resolution='custom',
opt_shape=opt_shape)
_, cent = centroid(im, im_blobs)
cent['frequency']= f[round(cent.y).astype(int)] # y values to frequency
# format rois to time-frequency
rois_out = format_rois(bbox, ts, f, fmt='tf')
# combine into a single df
aux_df = pd.concat([rois_out, shape, cent.frequency], axis=1)
# aux_df['fname'] = fname
features.append({'fname':fname, 'features': aux_df})
# Arranges the data into a dictionary
info_features = {'features': features,
'parameters_df': params,
'opt_shape': opt_shape,
'opt_spectro': opt_spec}
return info_features
rois_cr = rois_tf.loc[rois_tf.label=='CRER',]
rois_sp = rois_tf.loc[rois_tf.label=='SP',]
Sxx_power, ts, f, ext = spectrogram(s, fs)
Sxx_dB = power2dB(Sxx_power, db_range=90) + 96
# Visualize large vocalizations
rois_cr = format_features(rois_cr, ts, f)
overlay_rois(Sxx_dB, rois_cr, **{'extent':ext, 'vmin':0, 'vmax':80})
# Visualize short vocalizations
rois_sp = format_features(rois_sp, ts, f)
overlay_rois(Sxx_dB, rois_sp, **{'extent':ext, 'vmin':0, 'vmax':80})
# Compute an visualize features
shape_cr, params = shape_features(Sxx_dB, resolution='med', rois=rois_cr)
ax = plot_shape(shape_cr.mean(), params)
shape_sp, params = shape_features(Sxx_dB, resolution='med', rois=rois_sp)
ax = plot_shape(shape_sp.mean(), params)
######## Simple clustering with PCA
# join both shapes dataframe
features = shape_cr.append(shape_sp)
# Standardizing the dataset
X = features.filter(regex='shp*',axis='columns')
X_shape = X.values.shape
X = X.values.flatten()
X = X.reshape(-1, 1)
