malar
/
speech-scoring
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

implemented siamese pair tfrecord writer

master
Malar Kannan 2017-11-06 15:48:38 +05:30
parent c187fbe1ca
commit 046343680e
1 changed files with 47 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
speech_data.py
View File

@ -23,6 +23,15 @@ def get_siamese_pairs(groupF1, groupF2):
# return (random.sample(same,10), random.sample(diff,10))
return same[:10],diff[:10]
def siamese_pairs(rightGroup, wrongGroup):
group1 = [r for (i, r) in rightGroup.iterrows()]
group2 = [r for (i, r) in wrongGroup.iterrows()]
rightWrongPairs = [(g1, g2) for g2 in group2 for g1 in group1]
rightRightPairs = [i for i in itertools.combinations(group1, 2)]
random.shuffle(rightWrongPairs)
random.shuffle(rightRightPairs)
# return (random.sample(same,10), random.sample(diff,10))
return rightRightPairs[:10],rightWrongPairs[:10]
def append_zeros(spgr, max_samples):
return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
@ -92,7 +101,6 @@ def create_spectrogram_data(audio_group='audio'):
audio_samples.to_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
def create_spectrogram_tfrecords(audio_group='audio'):
# audio_group = 'story_words'
audio_samples = pd.read_csv( './outputs/' + audio_group + '.csv'
, names=['word','phonemes', 'voice', 'language', 'rate', 'variant', 'file']
, quoting=csv.QUOTE_NONE)
@ -101,9 +109,6 @@ def create_spectrogram_tfrecords(audio_group='audio'):
audio_samples['file_path'] = audio_samples.loc[:, 'file'].apply(lambda x: 'outputs/' + audio_group + '/' + x)
audio_samples['file_exists'] = apply_by_multiprocessing(audio_samples['file_path'], os.path.exists)
audio_samples = audio_samples[audio_samples['file_exists'] == True].reset_index()
# audio_samples['spectrogram'] = apply_by_multiprocessing(audio_samples['file_paths'],generate_aiff_spectrogram)#.apply(
# audio_samples['window_count'] = audio_samples.loc[:,'spectrogram'].apply(lambda x: x.shape[0])
# audio_samples.to_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
def _float_feature(value):
return tf.train.Feature(float_list=tf.train.FloatList(value=value))
@ -115,24 +120,41 @@ def create_spectrogram_tfrecords(audio_group='audio'):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))
writer = tf.python_io.TFRecordWriter('./outputs/' + audio_group + '.tfrecords')
audio_samples = audio_samples[:100]
for (i,sample) in audio_samples.iterrows():
spectrogram = generate_aiff_spectrogram(sample['file_path'])
spec_n = spectrogram.shape[0]
spec_w = spectrogram.shape[1]
spec = spectrogram.reshape(-1)
# audio_samples = audio_samples[:100]
for (w, word_group) in audio_samples.groupby(audio_samples['word']):
g = word_group.reset_index()
g['spectrogram'] = apply_by_multiprocessing(g['file_path'],generate_aiff_spectrogram)
sample_right = g.loc[audio_samples['variant'] == 'low']
sample_wrong = g.loc[audio_samples['variant'] == 'medium']
same, diff = siamese_pairs(sample_right, sample_wrong)
groups = [([0,1],same),([1,0],diff)]
for (output,group) in groups:
for sample1,sample2 in group:
spectro1,spectro2 = sample1['spectrogram'],sample2['spectrogram']
spec_n1,spec_n2 = spectro1.shape[0],spectro2.shape[0]
spec_w1,spec_w2 = spectro1.shape[1],spectro2.shape[1]
spec1,spec2 = spectro1.reshape(-1),spectro2.reshape(-1)
example = tf.train.Example(features=tf.train.Features(
feature={
'word': _bytes_feature([sample['word'].encode('utf-8')]),
'phoneme': _bytes_feature([sample['phonemes'].encode('utf-8')]),
'voice': _bytes_feature([sample['voice'].encode('utf-8')]),
'language': _bytes_feature([sample['language'].encode('utf-8')]),
'rate':_int64_feature([sample['rate']]),
'variant': _bytes_feature([sample['variant'].encode('utf-8')]),
'file': _bytes_feature([sample['file'].encode('utf-8')]),
'spec':_float_feature(spec),
'spec_n':_int64_feature([spec_n]),
'spec_w':_int64_feature([spec_w])
'word': _bytes_feature([w.encode('utf-8')]),
'phoneme1': _bytes_feature([sample1['phonemes'].encode('utf-8')]),
'phoneme2': _bytes_feature([sample2['phonemes'].encode('utf-8')]),
'voice1': _bytes_feature([sample1['voice'].encode('utf-8')]),
'voice2': _bytes_feature([sample2['voice'].encode('utf-8')]),
'language': _bytes_feature([sample1['language'].encode('utf-8')]),
'rate1':_int64_feature([sample1['rate']]),
'rate2':_int64_feature([sample2['rate']]),
'variant1': _bytes_feature([sample1['variant'].encode('utf-8')]),
'variant2': _bytes_feature([sample2['variant'].encode('utf-8')]),
'file1': _bytes_feature([sample1['file'].encode('utf-8')]),
'file2': _bytes_feature([sample2['file'].encode('utf-8')]),
'spec1':_float_feature(spec1),
'spec2':_float_feature(spec2),
'spec_n1':_int64_feature([spec_n1]),
'spec_w1':_int64_feature([spec_w1]),
'spec_n2':_int64_feature([spec_n2]),
'spec_w2':_int64_feature([spec_w2]),
'output':_int64_feature(output)
}
))
writer.write(example.SerializeToString())