114 lines
4.4 KiB
Python
114 lines
4.4 KiB
Python
import pandas as pd
|
|
import numpy as np
|
|
from spectro_gen import generate_aiff_spectrogram
|
|
from sklearn.model_selection import train_test_split
|
|
import itertools
|
|
import gc
|
|
|
|
|
|
def get_siamese_pairs(groupF1, groupF2):
|
|
group1 = [r for (i, r) in groupF1.iterrows()]
|
|
group2 = [r for (i, r) in groupF2.iterrows()]
|
|
f = [(g1, g2) for g2 in group2 for g1 in group1]
|
|
t = [i for i in itertools.combinations(group1, 2)
|
|
] + [i for i in itertools.combinations(group2, 2)]
|
|
return (t, f)
|
|
|
|
|
|
def create_X(sp, max_samples):
|
|
def append_zeros(spgr):
|
|
return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
|
|
'median')
|
|
|
|
l_sample = append_zeros(sp[0]['spectrogram'])
|
|
r_sample = append_zeros(sp[1]['spectrogram'])
|
|
return np.asarray([l_sample, r_sample])
|
|
|
|
|
|
def sunflower_pairs_data():
|
|
audio_samples = pd.read_csv(
|
|
'./outputs/audio.csv',
|
|
names=['word', 'voice', 'rate', 'variant', 'file'])
|
|
audio_samples = audio_samples.loc[audio_samples['word'] ==
|
|
'sunflowers'].reset_index(drop=True)
|
|
audio_samples.loc[:, 'spectrogram'] = audio_samples.loc[:, 'file'].apply(
|
|
lambda x: 'outputs/audio/' + x).apply(generate_aiff_spectrogram)
|
|
max_samples = audio_samples['spectrogram'].apply(
|
|
lambda x: x.shape[0]).max()
|
|
same_data, diff_data = [], []
|
|
for (w, g) in audio_samples.groupby(audio_samples['word']):
|
|
sample_norm = g.loc[audio_samples['variant'] == 'normal']
|
|
sample_phon = g.loc[audio_samples['variant'] == 'phoneme']
|
|
same, diff = get_siamese_pairs(sample_norm, sample_phon)
|
|
same_data.extend(same)
|
|
diff_data.extend(diff)
|
|
Y = np.hstack([np.ones(len(same_data)), np.zeros(len(diff_data))])
|
|
X_sample_pairs = same_data + diff_data
|
|
|
|
X_list = (create_X(sp, max_samples) for sp in X_sample_pairs)
|
|
X = np.vstack(X_list)
|
|
tr_pairs, te_pairs, tr_y, te_y = train_test_split(X, Y, test_size=0.1)
|
|
return train_test_split(X, Y, test_size=0.1)
|
|
|
|
|
|
def create_spectrogram_data(audio_group='audio'):
|
|
audio_samples = pd.read_csv(
|
|
'./outputs/' + audio_group + '.csv',
|
|
names=['word', 'voice', 'rate', 'variant', 'file'])
|
|
# audio_samples = audio_samples.loc[audio_samples['word'] ==
|
|
# 'sunflowers'].reset_index(drop=True)
|
|
audio_samples.loc[:, 'spectrogram'] = audio_samples.loc[:, 'file'].apply(
|
|
lambda x: 'outputs/' + audio_group + '/' + x).apply(
|
|
generate_aiff_spectrogram)
|
|
audio_samples.to_pickle('outputs/spectrogram.pkl')
|
|
|
|
|
|
def create_speech_pairs_data(audio_group='audio'):
|
|
audio_samples = pd.read_pickle('outputs/spectrogram.pkl')
|
|
max_samples = audio_samples['spectrogram'].apply(
|
|
lambda x: x.shape[0]).max()
|
|
# sample_size = audio_samples['spectrogram'][0].shape[1]
|
|
|
|
print('generating siamese speech pairs')
|
|
same_data, diff_data = [], []
|
|
for (w, g) in audio_samples.groupby(audio_samples['word']):
|
|
sample_norm = g.loc[audio_samples['variant'] == 'normal']
|
|
sample_phon = g.loc[audio_samples['variant'] == 'phoneme']
|
|
same, diff = get_siamese_pairs(sample_norm, sample_phon)
|
|
same_data.extend([create_X(s, max_samples) for s in same[:10]])
|
|
diff_data.extend([create_X(d, max_samples) for d in diff[:10]])
|
|
print('creating all speech pairs')
|
|
Y = np.hstack([np.ones(len(same_data)), np.zeros(len(diff_data))])
|
|
print('casting as array speech pairs')
|
|
X = np.asarray(same_data + diff_data)
|
|
print('pickling X/Y')
|
|
np.save('outputs/X.npy', X)
|
|
np.save('outputs/Y.npy', Y)
|
|
del same_data
|
|
del diff_data
|
|
gc.collect()
|
|
print('train/test splitting speech pairs')
|
|
tr_pairs, te_pairs, tr_y, te_y = train_test_split(X, Y, test_size=0.1)
|
|
print('pickling train/test')
|
|
np.save('outputs/tr_pairs.npy', tr_pairs)
|
|
np.save('outputs/te_pairs.npy', te_pairs)
|
|
np.save('outputs/tr_y.npy', tr_y)
|
|
np.save('outputs/te_y.npy', te_y)
|
|
|
|
|
|
def speech_model_data():
|
|
tr_pairs = np.load('outputs/tr_pairs.npy') / 255.0
|
|
te_pairs = np.load('outputs/te_pairs.npy') / 255.0
|
|
tr_pairs[tr_pairs < 0] = 0
|
|
te_pairs[te_pairs < 0] = 0
|
|
tr_y = np.load('outputs/tr_y.npy')
|
|
te_y = np.load('outputs/te_y.npy')
|
|
return tr_pairs, te_pairs, tr_y, te_y
|
|
|
|
|
|
if __name__ == '__main__':
|
|
# sunflower_pairs_data()
|
|
# create_spectrogram_data()
|
|
create_speech_pairs_data()
|
|
# print(speech_model_data())
|