def audio_mixup_augmentation(paths_list, cleared_audio_dir, alpha, num_classes,
sample_rate, num_coeffs, num_timesteps,
num_to_mix):
print('Started data augmentation using audio mixup...')
time_start = datetime.datetime.now()
mixup_features = []
mixup_labels = []
paths_list = list(dict.fromkeys(paths_list)) # remove all duplicates
num_records = len(paths_list)
np.random.shuffle(paths_list)
audio, labels = load_cleared_audio(cleared_audio_dir, paths_list)
labels = to_categorical(labels, num_classes)
print('Augmentation:')
num_same_class = 0
for idx in range(num_records - num_to_mix + 1):
for mix_idx in range(num_to_mix - 1):
lam = np.random.beta(a=alpha, b=alpha)
audio_size = len(audio[idx]) if len(audio[idx]) < len(
audio[idx + mix_idx + 1]) else len(audio[idx + mix_idx + 1])
mixed_audio = lam * audio[idx][:audio_size] + (
1 - lam) * audio[idx + mix_idx + 1][:audio_size]
mixed_lab = lam * labels[idx] + (1 - lam) * labels[idx + mix_idx +
1]
extractor = FeatureExtractor(mixed_audio, sample_rate, num_coeffs)
mixed_feat = extractor.extract_log_mel_filterbank_energies()
timestep_samples = compose_samples_with_timesteps(
mixed_feat, num_timesteps)
mixup_features.extend(timestep_samples)
stretched_label = np.tile(mixed_lab,
(timestep_samples.shape[0], 1))
mixup_labels.extend(stretched_label)
if np.array_equal(labels[idx], labels[idx + mix_idx + 1]):
num_same_class += 1
print_progress_bar(idx,
num_records - num_to_mix,
prefix='{}/{}'.format(idx + 1,
num_records - num_to_mix + 1),
suffix='complete')
time_end = datetime.datetime.now()
print('Elapsed time: ', time_end - time_start)
ratio = num_same_class / (num_to_mix * num_records)
print('%.2f%% are mixed from the same class,\
\n%.2f%% are mixed from different classes' % (ratio, 1 - ratio))
return np.array(mixup_features), np.array(mixup_labels)
def extract_mel_filterbank_energies(data, sample_rate, num_coeffs):
extractor = FeatureExtractor(data_array=data,
sample_rate=sample_rate,
num_of_coeffs=num_coeffs)
features = extractor.extract_log_mel_filterbank_energies()
return features
