def test_spectrogram_256():
n_fft = 256
sample_rate = 250000
samples = sample_rate * 1
spectrogram = Audio.load_tone(
20, samples, 'sine', sample_rate).get_spectrogram_array(n_fft=n_fft)
spectrogram = to_uint8(spectrogram)
assert spectrogram.shape[0] == (n_fft / 2) + 1, 'spectrogram height'
assert spectrogram.shape[1] == int(
samples * 2 / n_fft) + 1, 'spectrogram width'
assert np.array([
all(0 <= y <= 255 for y in x) for x in spectrogram
]).all(), f'range check {spectrogram} failed. Expected all values 0-255'
def test_mel_spectrogram_256():
n_fft = 256
hop_length = 128
n_mels = 60
sample_rate = 384000
samples = sample_rate * 1
spectrogram = Audio.load_tone(20, samples, 'sine',
sample_rate).get_mel_spectrogram_array(
n_fft=n_fft,
n_mels=n_mels,
hop_length=hop_length)
spectrogram = to_uint8(spectrogram)
print(spectrogram.shape)
assert spectrogram.shape[0] == n_mels, 'spectrogram height'
assert spectrogram.shape[1] == int(
samples * 2 / n_fft) + 1, 'spectrogram width'
assert np.array([
all(0 <= y <= 255 for y in x) for x in spectrogram
]).all(), f'range check {spectrogram} failed. Expected all values 0-255'
def generate_jonnor_mnist(self,
preprocessed,
train_folds=[1, 2, 3, 4],
test_folds=[5]):
x_train = []
y_train = []
s_train = []
x_test = []
y_test = []
s_test = []
for fold, filename, target, category, take, spectrogram, settings in preprocessed:
d = to_uint8(spectrogram).flatten()
if fold in train_folds:
x_train.append(d)
y_train.append(target)
s_train.append(spectrogram.shape)
elif fold in test_folds:
x_test.append(d)
y_test.append(target)
s_test.append(spectrogram.shape)
return x_train, y_train, s_train, x_test, y_test, s_test
def to_mnist(self,
train_folds=[],
test_folds=[],
cache_path: str = None,
n_fft: int = 1024,
hop_length: int = None,
flatten=True):
# (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
# x_train = arrays of data in [0:255] range
# y_train = label (class)
#Setup Caching
cache_training_fold: str = "".join([str(x) for x in train_folds])
cache_testing_fold: str = "".join([str(x) for x in test_folds])
cache_file = f'mnist_{str(n_fft)}_{str(hop_length)}_{str(int(flatten))}_{cache_training_fold}_{cache_testing_fold}.pkl'
cache_filename = os.path.join(cache_path, cache_file)
if cache_path is not None:
print(f'Caching: {cache_filename}')
if os.path.exists(cache_filename):
with open(cache_filename, 'rb') as f:
return pickle.load(f)
if hop_length is None:
hop_length = int(n_fft / 4)
all_records = list(
zip(self.get_folds(), self.get_filenames(), self.get_targets()))
x_train = []
y_train = []
s_train = []
x_test = []
y_test = []
s_test = []
for folds, x, y, s in tqdm([(train_folds, x_train, y_train, s_train),
(test_folds, x_test, y_test, s_test)],
desc='Fold'):
for fold, filename, target in tqdm(
[r for r in all_records if int(r[0]) in folds], desc='File'):
audio = self.get_audio(filename)
if audio is None:
continue
# Get a Numpy array of the spectrogram
spectrogram = audio.get_spectrogram_array(
n_fft=n_fft, hop_length=hop_length, mode='dB')
# Normalise - Subtract mean and divide by Standard deviation
spectrogram = to_uint8(spectrogram)
# Flatten spectrogram if required and store
if flatten:
x.append(spectrogram.flatten())
else:
x.append(spectrogram)
y.append(target)
s.append(spectrogram.shape)
if cache_path is not None:
with open(cache_filename, 'wb') as f:
pickle.dump(
(x_train, y_train, s_train, x_test, y_test, s_test), f)
return x_train, y_train, s_train, x_test, y_test, s_test