def test1(self):
audio, output_audio = make_sine_waves()
audio_tensor = tf.convert_to_tensor(audio, dtype=tf.float32)
output_audio_tensor = tf.convert_to_tensor(output_audio,
dtype=tf.float32)
input_batch = mu_law_encode(audio_tensor, 256)
output_batch = mu_law_encode(output_audio_tensor, 256)
encoded = self.net._one_hot(input_batch)
output_encoded = self.net._one_hot(output_batch)
shifted = tf.slice(output_encoded, [0, 1, 0],
[-1, tf.shape(output_encoded)[1] - 1, -1])
# shifted = tf.pad(shifted, [[0, 0], [0, 1], [0, 0]])
raw_output = self.net._create_network(encoded)
out = tf.reshape(raw_output, [-1, self.net.quantization_channels])
# Cast to float64 to avoid bug in TensorFlow
proba = tf.cast(tf.nn.softmax(tf.cast(out, tf.float64)), tf.float32)
last = tf.slice(proba, [tf.shape(proba)[0] - 1, 0],
[1, self.net.quantization_channels])
lasted = tf.reshape(last, [-1])
# shifted = tf.pad(shifted, [[0, 0], [0, 1], [0, 0]])
# slice = tf.reshape(shifted, [-1, self.net.quantization_channels])
with self.test_session() as sess:
sess.run(tf.initialize_all_variables())
print(sess.run(out).shape)
print(sess.run(proba)[1])
print(sess.run(proba)[0])
print(sess.run(last).shape)
print(sess.run(lasted).shape)
def load_audio_not_one_hot(
filename,
sample_rate=get_model_params('SAMPLE_RATE'),
quantization_channels=get_model_params('QUANTIZATION_CHANNELS'),
batch_size=get_model_params('BATCH_SIZE')):
audio = load_wav(filename, sample_rate)
quantized = mu_law_encode(audio, quantization_channels)
return quantized
def create_seed(filename, sample_rate, quantization_channels, window_size):
audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
quantized = mu_law_encode(audio, quantization_channels)
cut_index = tf.cond(
tf.size(quantized) < tf.constant(window_size),
lambda: tf.size(quantized), lambda: tf.constant(window_size))
return quantized[:cut_index]
def testEncodeDecode(self):
x = np.linspace(-1, 1, 1000).astype(np.float32)
channels = 256
# Test whether decoded signal is roughly equal to
# what was encoded before
with self.test_session() as sess:
encoded = mu_law_encode(x, channels)
x1 = sess.run(mu_law_decode(encoded, channels))
self.assertAllClose(x, x1, rtol=1e-1, atol=0.05)
# Make sure that re-encoding leaves the waveform invariant
with self.test_session() as sess:
encoded = mu_law_encode(x1, channels)
x2 = sess.run(mu_law_decode(encoded, channels))
self.assertAllClose(x1, x2)
def testEncodeDecode(self):
x = np.linspace(-1, 1, 1000).astype(np.float32)
channels = 256
# Test whether decoded signal is roughly equal to
# what was encoded before
with self.test_session() as sess:
encoded = mu_law_encode(x, channels)
x1 = sess.run(mu_law_decode(encoded, channels))
self.assertAllClose(x, x1, rtol=1e-1, atol=0.05)
# Make sure that re-encoding leaves the waveform invariant
with self.test_session() as sess:
encoded = mu_law_encode(x1, channels)
x2 = sess.run(mu_law_decode(encoded, channels))
self.assertAllClose(x1, x2)
def testEndToEndTraining(self):
audio = make_sine_waves()
np.random.seed(42)
# if self.generate:
# librosa.output.write_wav('/tmp/sine_train.wav', audio,
# SAMPLE_RATE_HZ)
# power_spectrum = np.abs(np.fft.fft(audio))**2
# freqs = np.fft.fftfreq(audio.size, SAMPLE_PERIOD_SECS)
# indices = np.argsort(freqs)
# indices = [index for index in indices if freqs[index] >= 0 and
# freqs[index] <= 500.0]
# plt.plot(freqs[indices], power_spectrum[indices])
# plt.show()
audio_tensor = tf.convert_to_tensor(audio, dtype=tf.float32)
encode_output = mu_law_encode(audio_tensor, QUANTIZATION_CHANNELS)
loss = self.net.loss(encode_output)
optimizer = optimizer_factory[self.optimizer_type](
learning_rate=self.learning_rate, momentum=self.momentum)
trainable = tf.trainable_variables()
optim = optimizer.minimize(loss, var_list=trainable)
init = tf.initialize_all_variables()
generated_waveform = None
max_allowed_loss = 0.1
loss_val = max_allowed_loss
initial_loss = None
with self.test_session() as sess:
sess.run(init)
initial_loss = sess.run(loss)
for i in range(TRAIN_ITERATIONS):
loss_val, _ = sess.run([loss, optim])
# if i % 10 == 0:
# print("i: %d loss: %f" % (i, loss_val))
# Sanity check the initial loss was larger.
self.assertGreater(initial_loss, max_allowed_loss)
# Loss after training should be small.
self.assertLess(loss_val, max_allowed_loss)
# Loss should be at least two orders of magnitude better
# than before training.
self.assertLess(loss_val / initial_loss, 0.01)
# saver = tf.train.Saver(var_list=tf.trainable_variables())
# saver.save(sess, '/tmp/sine_test_model.ckpt', global_step=i)
if self.generate:
# Check non-incremental generation
generated_waveform = generate_waveform(sess, self.net, False)
check_waveform(self.assertGreater, generated_waveform)
# Check incremental generation
generated_waveform = generate_waveform(sess, self.net, True)
check_waveform(self.assertGreater, generated_waveform)
def testEncodeNegativeChannelSize(self):
np.random.seed(1944) # For repeatability of test.
channels = -256
number_of_samples = 1024
x = np.zeros(number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
self.assertRaises(TypeError, sess.run(mu_law_encode(x, channels)))
def testEncodeNegativeChannelSize(self):
np.random.seed(1944) # For repeatability of test.
channels = -256
number_of_samples = 1024
x = np.zeros(number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
self.assertRaises(TypeError, sess.run(mu_law_encode(x, channels)))
def testEncodeUniformRandomNoise(self):
np.random.seed(42) # For repeatability of test.
channels = 256
number_of_samples = 2048
x = np.random.uniform(-1, 1, number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def testEncodeZeros(self):
np.random.seed(1944) # For repeatability of test.
channels = 256
number_of_samples = 1024
x = np.zeros(number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def testEncodePrecomputed(self):
channels = 256
number_of_samples = 10
x = np.array([-1.0, 1.0, 0.6, -0.25, 0.01,
0.33, -0.9999, 0.42, 0.1, -0.45]).astype(np.float32)
encoded_manual = np.array([0, 255, 243, 32, 157,
230, 0, 235, 203, 18]).astype(np.int32)
with self.test_session() as sess:
encoded = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(encoded_manual, encoded)
def testEncodeUniformRandomNoise(self):
np.random.seed(42) # For repeatability of test.
channels = 256
number_of_samples = 2048
x = np.random.uniform(-1, 1, number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def testEncodeZeros(self):
np.random.seed(1944) # For repeatability of test.
channels = 256
number_of_samples = 1024
x = np.zeros(number_of_samples).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def testEncodeRamp(self):
np.random.seed(1944) # For repeatability of test.
channels = 256
number_of_samples = 1024
number_of_steps = 2.0 / number_of_samples
x = np.arange(-1.0, 1.0, number_of_steps).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def create_seed(filename,
sample_rate,
quantization_channels,
window_size=WINDOW):
audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
audio = audio_reader.trim_silence(audio)
quantized = mu_law_encode(audio, quantization_channels)
cut_index = tf.cond(tf.size(quantized) < tf.constant(window_size),
lambda: tf.size(quantized),
lambda: tf.constant(window_size))
return quantized[:cut_index]
def create_seed(waveform,
sample_rate,
quantization_channels,
window_size,
silence_threshold=SILENCE_THRESHOLD):
#audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
#audio = audio_reader.trim_silence(audio, silence_threshold)
quantized = mu_law_encode(waveform, quantization_channels)
# explaning lambda function this is the condition if true return size of quantized else return const size (varies)
cut_index = tf.cond(tf.size(quantized) < tf.constant(window_size), lambda: tf.size(quantized), lambda: tf.constant(window_size))
return quantized[:cut_index]
def testEncodePrecomputed(self):
channels = 256
number_of_samples = 10
x = np.array(
[-1.0, 1.0, 0.6, -0.25, 0.01, 0.33, -0.9999, 0.42, 0.1,
-0.45]).astype(np.float32)
encoded_manual = np.array([0, 255, 243, 32, 157, 230, 0, 235, 203,
18]).astype(np.int32)
with self.test_session() as sess:
encoded = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(encoded_manual, encoded)
def create_seed(filename,
sample_rate,
quantization_channels,
window_size,
silence_threshold=SILENCE_THRESHOLD):
audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
#audio = audio_reader.trim_silence(audio, silence_threshold)
quantized = mu_law_encode(audio, quantization_channels)
cut_index = tf.cond(tf.size(quantized) < tf.constant(window_size),
lambda: tf.size(quantized),
lambda: tf.constant(window_size))
return quantized[:cut_index]
def testEncodeRamp(self):
np.random.seed(1944) # For repeatability of test.
channels = 256
number_of_samples = 1024
number_of_steps = 2.0 / number_of_samples
x = np.arange(-1.0, 1.0, number_of_steps).astype(np.float32)
manual_encode = manual_mu_law_encode(x, channels)
with self.test_session() as sess:
encode = sess.run(mu_law_encode(x, channels))
self.assertAllEqual(manual_encode, encode)
def testDecodeEncode(self):
# generate every possible quantized level.
x = np.array(range(QUANT_LEVELS), dtype=np.int)
# Encoded then decode every value.
with self.test_session() as sess:
# Decode into floating-point scalar.
decoded = mu_law_decode(x, QUANT_LEVELS)
# Encode back into an integer quantization level.
encoded = mu_law_encode(decoded, QUANT_LEVELS)
round_tripped = sess.run(encoded)
# decoding then encoding every level should produce what we started
# with.
self.assertAllEqual(x, round_tripped)
def testDecodeEncode(self):
# generate every possible quantized level.
x = np.array(range(QUANT_LEVELS), dtype=np.int)
# Encoded then decode every value.
with self.test_session() as sess:
# Decode into floating-point scalar.
decoded = mu_law_decode(x, QUANT_LEVELS)
# Encode back into an integer quantization level.
encoded = mu_law_encode(decoded, QUANT_LEVELS)
round_tripped = sess.run(encoded)
# decoding then encoding every level should produce what we started
# with.
self.assertAllEqual(x, round_tripped)
def testEncodeDecodeShift(self):
x = np.linspace(-1, 1, 1000).astype(np.float32)
with self.test_session() as sess:
encoded = mu_law_encode(x, QUANT_LEVELS)
decoded = mu_law_decode(encoded, QUANT_LEVELS)
roundtripped = sess.run(decoded)
# Detect non-unity scaling and non-zero shift in the roundtripped
# signal by asserting that slope = 1 and y-intercept = 0 of line fit to
# roundtripped vs x values.
coeffs = np.polyfit(x, roundtripped, 1)
slope = coeffs[0]
y_intercept = coeffs[1]
EPSILON = 1e-4
self.assertNear(slope, 1.0, EPSILON)
self.assertNear(y_intercept, 0.0, EPSILON)
def create_seed(filename,sample_rate,quantization_channels,window_size,scalar_input):
# seed의 앞부분만 사용한다.
seed_audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
seed_audio = audio.trim_silence(seed_audio, default_hparams)
if scalar_input:
if len(seed_audio) < window_size:
return seed_audio
else: return seed_audio[:window_size]
else:
quantized = mu_law_encode(seed_audio, quantization_channels)
# 짧으면 짧은 대로 return하는데, padding이라도 해야되지 않나???
cut_index = tf.cond(tf.size(quantized) < tf.constant(window_size), lambda: tf.size(quantized), lambda: tf.constant(window_size))
return quantized[:cut_index]
def testEncodeDecodeShift(self):
x = np.linspace(-1, 1, 1000).astype(np.float32)
with self.test_session() as sess:
encoded = mu_law_encode(x, QUANT_LEVELS)
decoded = mu_law_decode(encoded, QUANT_LEVELS)
roundtripped = sess.run(decoded)
# Detect non-unity scaling and non-zero shift in the roundtripped
# signal by asserting that slope = 1 and y-intercept = 0 of line fit to
# roundtripped vs x values.
coeffs = np.polyfit(x, roundtripped, 1)
slope = coeffs[0]
y_intercept = coeffs[1]
EPSILON = 1e-4
self.assertNear(slope, 1.0, EPSILON)
self.assertNear(y_intercept, 0.0, EPSILON)
def create_seed(filename,
sample_rate,
quantization_channels,
window_size,
scalar_input,
silence_threshold=SILENCE_THRESHOLD):
audio, _ = librosa.load(filename, sr=sample_rate, mono=True)
audio = audio_reader.trim_silence(audio, silence_threshold)
if scalar_input:
if len(audio) < window_size:
return audio
else:
return audio[:window_size]
else:
quantized = mu_law_encode(audio, quantization_channels)
# 짧으면 짧은 대로 return하는데, padding이라도 해야되지 않나???
cut_index = tf.cond(
tf.size(quantized) < tf.constant(window_size),
lambda: tf.size(quantized), lambda: tf.constant(window_size))
return quantized[:cut_index]
x for x in os.listdir(DIRS['SONGS']) if x.endswith('.wav')
][0])
wav_fname_new = wav_fname.replace('.wav', '_after.wav')
# low raw audio
# In[14]:
audio, _ = librosa.load(wav_fname, sr=M_PARAMS['SAMPLE_RATE'], mono=True)
audio[1000:1050]
# encode it to 8 bit amplitude
# In[15]:
quantized = mu_law_encode(audio, M_PARAMS['QUANTISATION_CHANNELS'])
quantized[1000:1050].eval(session=sess)
# get RNN input
# In[16]:
quantized_oh = _one_hot(quantized)
quantized_oh[0][1000:1020].eval(session=sess)
# let RNN out be exact RNN input (for test)
#
# turn it back to 8 bit signal
# In[17]:
def testEncodeIsSurjective(self):
x = np.linspace(-1, 1, 10000).astype(np.float32)
channels = 123
with self.test_session() as sess:
encoded = sess.run(mu_law_encode(x, channels))
self.assertEqual(len(np.unique(encoded)), channels)
def testEncodeIsSurjective(self):
x = np.linspace(-1, 1, 10000).astype(np.float32)
channels = 123
with self.test_session() as sess:
encoded = sess.run(mu_law_encode(x, channels))
self.assertEqual(len(np.unique(encoded)), channels)
