def _compare(self, signal, frame_length, frame_step, fft_length):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)) as sess:
actual_stft = spectral_ops.stft(
signal, frame_length, frame_step, fft_length, pad_end=False)
signal_ph = array_ops.placeholder(dtype=dtypes.as_dtype(signal.dtype))
actual_stft_from_ph = spectral_ops.stft(
signal_ph, frame_length, frame_step, fft_length, pad_end=False)
actual_inverse_stft = spectral_ops.inverse_stft(
actual_stft, frame_length, frame_step, fft_length)
actual_stft, actual_stft_from_ph, actual_inverse_stft = sess.run(
[actual_stft, actual_stft_from_ph, actual_inverse_stft],
feed_dict={signal_ph: signal})
actual_stft_ph = array_ops.placeholder(dtype=actual_stft.dtype)
actual_inverse_stft_from_ph = sess.run(
spectral_ops.inverse_stft(
actual_stft_ph, frame_length, frame_step, fft_length),
feed_dict={actual_stft_ph: actual_stft})
# Confirm that there is no difference in output when shape/rank is fully
# unknown or known.
self.assertAllClose(actual_stft, actual_stft_from_ph)
self.assertAllClose(actual_inverse_stft, actual_inverse_stft_from_ph)
expected_stft = SpectralOpsTest._np_stft(
signal, fft_length, frame_step, frame_length)
self.assertAllClose(expected_stft, actual_stft, 1e-4, 1e-4)
expected_inverse_stft = SpectralOpsTest._np_inverse_stft(
expected_stft, fft_length, frame_step, frame_length)
self.assertAllClose(
expected_inverse_stft, actual_inverse_stft, 1e-4, 1e-4)
def test_shapes(self):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)):
signal = np.zeros((512,)).astype(np.float32)
# If fft_length is not provided, the smallest enclosing power of 2 of
# frame_length (8) is used.
stft = spectral_ops.stft(signal, frame_length=7, frame_step=8,
pad_end=True)
self.assertAllEqual([64, 5], stft.shape.as_list())
self.assertAllEqual([64, 5], stft.eval().shape)
stft = spectral_ops.stft(signal, frame_length=8, frame_step=8,
pad_end=True)
self.assertAllEqual([64, 5], stft.shape.as_list())
self.assertAllEqual([64, 5], stft.eval().shape)
stft = spectral_ops.stft(signal, frame_length=8, frame_step=8,
fft_length=16, pad_end=True)
self.assertAllEqual([64, 9], stft.shape.as_list())
self.assertAllEqual([64, 9], stft.eval().shape)
stft = np.zeros((32, 9)).astype(np.complex64)
inverse_stft = spectral_ops.inverse_stft(stft, frame_length=8,
fft_length=16, frame_step=8)
expected_length = (stft.shape[0] - 1) * 8 + 8
self.assertAllEqual([None], inverse_stft.shape.as_list())
self.assertAllEqual([expected_length], inverse_stft.eval().shape)
def test_shapes(self):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)):
signal = np.zeros((512,)).astype(np.float32)
# If fft_length is not provided, the smallest enclosing power of 2 of
# frame_length (8) is used.
stft = spectral_ops.stft(signal, frame_length=7, frame_step=8,
pad_end=True)
self.assertAllEqual([64, 5], stft.shape.as_list())
self.assertAllEqual([64, 5], stft.eval().shape)
stft = spectral_ops.stft(signal, frame_length=8, frame_step=8,
pad_end=True)
self.assertAllEqual([64, 5], stft.shape.as_list())
self.assertAllEqual([64, 5], stft.eval().shape)
stft = spectral_ops.stft(signal, frame_length=8, frame_step=8,
fft_length=16, pad_end=True)
self.assertAllEqual([64, 9], stft.shape.as_list())
self.assertAllEqual([64, 9], stft.eval().shape)
stft = spectral_ops.stft(signal, frame_length=16, frame_step=8,
fft_length=8, pad_end=True)
self.assertAllEqual([64, 5], stft.shape.as_list())
self.assertAllEqual([64, 5], stft.eval().shape)
stft = np.zeros((32, 9)).astype(np.complex64)
inverse_stft = spectral_ops.inverse_stft(stft, frame_length=8,
fft_length=16, frame_step=8)
expected_length = (stft.shape[0] - 1) * 8 + 8
self.assertAllEqual([None], inverse_stft.shape.as_list())
self.assertAllEqual([expected_length], inverse_stft.eval().shape)
def testContribSignalSTFT(self):
ws = 512
hs = 128
dims = (ws * 20,)
shape = BATCH_DIMS + dims
data = np.arange(np.prod(shape)) / np.prod(dims)
np.random.seed(123)
np.random.shuffle(data)
data = np.reshape(data.astype(np.float32), shape)
window = sps.get_window("hann", ws)
expected = sps.stft(
data, nperseg=ws, noverlap=ws - hs, boundary=None, window=window)[2]
expected = np.swapaxes(expected, -1, -2)
expected *= window.sum() # scipy divides by window sum
with self.test_session() as sess:
with self.test_scope():
ph = array_ops.placeholder(
dtypes.as_dtype(data.dtype), shape=data.shape)
out = signal.stft(ph, ws, hs)
grad = gradients_impl.gradients(out, ph,
grad_ys=array_ops.ones_like(out))
# For gradients, we simply verify that they compile & execute.
value, _ = sess.run([out, grad], {ph: data})
self.assertAllClose(expected, value, rtol=RTOL, atol=ATOL)
def testContribSignalSTFT(self):
ws = 512
hs = 128
dims = (ws * 20, )
shape = BATCH_DIMS + dims
data = np.arange(np.prod(shape)) / np.prod(dims)
np.random.seed(123)
np.random.shuffle(data)
data = np.reshape(data.astype(np.float32), shape)
window = sps.get_window("hann", ws)
expected = sps.stft(data,
nperseg=ws,
noverlap=ws - hs,
boundary=None,
window=window)[2]
expected = np.swapaxes(expected, -1, -2)
expected *= window.sum() # scipy divides by window sum
with self.test_session() as sess:
with self.test_scope():
ph = array_ops.placeholder(dtypes.as_dtype(data.dtype),
shape=data.shape)
out = signal.stft(ph, ws, hs)
grad = gradients_impl.gradients(
out, ph, grad_ys=array_ops.ones_like(out))
# For gradients, we simply verify that they compile & execute.
value, _ = sess.run([out, grad], {ph: data})
self.assertAllClose(expected, value, rtol=RTOL, atol=ATOL)
def test_gradients_numerical(self):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)):
# Tuples of (signal_length, frame_length, frame_step, fft_length,
# stft_bound, inverse_stft_bound).
# TODO(rjryan): Investigate why STFT gradient error is so high.
test_configs = [
(64, 16, 8, 16),
(64, 16, 16, 16),
(64, 16, 7, 16),
(64, 7, 4, 9),
(29, 5, 1, 10),
]
for (signal_length, frame_length, frame_step, fft_length) in test_configs:
signal_shape = [signal_length]
signal = random_ops.random_uniform(signal_shape)
stft_shape = [max(0, 1 + (signal_length - frame_length) // frame_step),
fft_length // 2 + 1]
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length,
pad_end=False)
inverse_stft_shape = [(stft_shape[0] - 1) * frame_step + frame_length]
inverse_stft = spectral_ops.inverse_stft(stft, frame_length, frame_step,
fft_length)
stft_error = test.compute_gradient_error(signal, [signal_length],
stft, stft_shape)
inverse_stft_error = test.compute_gradient_error(
stft, stft_shape, inverse_stft, inverse_stft_shape)
self.assertLess(stft_error, 2e-3)
self.assertLess(inverse_stft_error, 5e-4)
def _compute_stft_gradient(signal, frame_length=32, frame_step=16,
fft_length=32):
"""Computes the gradient of the STFT with respect to `signal`."""
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length)
magnitude_stft = math_ops.abs(stft)
loss = math_ops.reduce_sum(magnitude_stft)
return gradients_impl.gradients([loss], [signal])[0]
def _compare_round_trip(self, signal, frame_length, frame_step,
fft_length):
with spectral_ops_test_util.fft_kernel_label_map(), (self.test_session(
use_gpu=True)) as sess:
stft = spectral_ops.stft(signal,
frame_length,
frame_step,
fft_length,
pad_end=False)
inverse_stft = spectral_ops.inverse_stft(stft, frame_length,
frame_step, fft_length)
signal, inverse_stft = sess.run([signal, inverse_stft])
# Since the shapes can differ due to padding, pad both signals to the max
# of their lengths.
max_length = max(signal.shape[0], inverse_stft.shape[0])
signal = np.pad(signal, (0, max_length - signal.shape[0]),
"constant")
inverse_stft = np.pad(inverse_stft,
(0, max_length - inverse_stft.shape[0]),
"constant")
# Ignore the frame_length samples at either edge.
start = frame_length
end = signal.shape[0] - frame_length
ratio = signal[start:end] / inverse_stft[start:end]
# Check that the inverse and original signal are equal up to a constant
# factor.
self.assertLess(np.var(ratio), 2e-5)
def test_gradients_numerical(self):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)):
# Tuples of (signal_length, frame_length, frame_step, fft_length,
# stft_bound, inverse_stft_bound).
# TODO(rjryan): Investigate why STFT gradient error is so high.
test_configs = [
(64, 16, 8, 16),
(64, 16, 16, 16),
(64, 16, 7, 16),
(64, 7, 4, 9),
(29, 5, 1, 10),
]
for (signal_length, frame_length, frame_step, fft_length) in test_configs:
signal_shape = [signal_length]
signal = random_ops.random_uniform(signal_shape)
stft_shape = [max(0, 1 + (signal_length - frame_length) // frame_step),
fft_length // 2 + 1]
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length,
pad_end=False)
inverse_stft_shape = [(stft_shape[0] - 1) * frame_step + frame_length]
inverse_stft = spectral_ops.inverse_stft(stft, frame_length, frame_step,
fft_length)
stft_error = test.compute_gradient_error(signal, [signal_length],
stft, stft_shape)
inverse_stft_error = test.compute_gradient_error(
stft, stft_shape, inverse_stft, inverse_stft_shape)
self.assertLess(stft_error, 2e-3)
self.assertLess(inverse_stft_error, 5e-4)
def _compute_stft_gradient(signal, frame_length=32, frame_step=16,
fft_length=32):
"""Computes the gradient of the STFT with respect to `signal`."""
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length)
magnitude_stft = math_ops.abs(stft)
loss = math_ops.reduce_sum(magnitude_stft)
return gradients_impl.gradients([loss], [signal])[0]
def test_stft_round_trip(self):
# Tuples of (signal_length, frame_length, frame_step, fft_length,
# threshold, corrected_threshold).
test_configs = [
# 87.5% overlap.
(4096, 256, 32, 256, 1e-5, 1e-6),
# 75% overlap.
(4096, 256, 64, 256, 1e-5, 1e-6),
# Odd frame hop.
(4096, 128, 25, 128, 1e-3, 1e-6),
# Odd frame length.
(4096, 127, 32, 128, 1e-3, 1e-6),
# 50% overlap.
(4096, 128, 64, 128, 0.40, 1e-6),
]
for (signal_length, frame_length, frame_step, fft_length, threshold,
corrected_threshold) in test_configs:
# Generate a random white Gaussian signal.
signal = random_ops.random_normal([signal_length])
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)) as sess:
stft = spectral_ops.stft(signal,
frame_length,
frame_step,
fft_length,
pad_end=False)
inverse_stft = spectral_ops.inverse_stft(
stft, frame_length, frame_step, fft_length)
inverse_stft_corrected = spectral_ops.inverse_stft(
stft,
frame_length,
frame_step,
fft_length,
window_fn=spectral_ops.inverse_stft_window_fn(frame_step))
signal, inverse_stft, inverse_stft_corrected = sess.run(
[signal, inverse_stft, inverse_stft_corrected])
# Truncate signal to the size of inverse stft.
signal = signal[:inverse_stft.shape[0]]
# Ignore the frame_length samples at either edge.
signal = signal[frame_length:-frame_length]
inverse_stft = inverse_stft[frame_length:-frame_length]
inverse_stft_corrected = inverse_stft_corrected[
frame_length:-frame_length]
# Check that the inverse and original signal are close up to a scale
# factor.
inverse_stft_scaled = inverse_stft / np.mean(
np.abs(inverse_stft))
signal_scaled = signal / np.mean(np.abs(signal))
self.assertLess(np.std(inverse_stft_scaled - signal_scaled),
threshold)
# Check that the inverse with correction and original signal are close.
self.assertLess(np.std(inverse_stft_corrected - signal),
corrected_threshold)
def test_stft_round_trip(self):
# Tuples of (signal_length, frame_length, frame_step, fft_length,
# threshold, corrected_threshold).
test_configs = [
# 87.5% overlap.
(4096, 256, 32, 256, 1e-5, 1e-6),
# 75% overlap.
(4096, 256, 64, 256, 1e-5, 1e-6),
# Odd frame hop.
(4096, 128, 25, 128, 1e-3, 1e-6),
# Odd frame length.
(4096, 127, 32, 128, 1e-3, 1e-6),
# 50% overlap.
(4096, 128, 64, 128, 0.40, 1e-6),
]
for (signal_length, frame_length, frame_step, fft_length, threshold,
corrected_threshold) in test_configs:
# Generate a random white Gaussian signal.
signal = random_ops.random_normal([signal_length])
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)) as sess:
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length,
pad_end=False)
inverse_stft = spectral_ops.inverse_stft(stft, frame_length, frame_step,
fft_length)
inverse_stft_corrected = spectral_ops.inverse_stft(
stft, frame_length, frame_step, fft_length,
window_fn=spectral_ops.inverse_stft_window_fn(frame_step))
signal, inverse_stft, inverse_stft_corrected = sess.run(
[signal, inverse_stft, inverse_stft_corrected])
# Truncate signal to the size of inverse stft.
signal = signal[:inverse_stft.shape[0]]
# Ignore the frame_length samples at either edge.
signal = signal[frame_length:-frame_length]
inverse_stft = inverse_stft[frame_length:-frame_length]
inverse_stft_corrected = inverse_stft_corrected[
frame_length:-frame_length]
# Check that the inverse and original signal are close up to a scale
# factor.
inverse_stft_scaled = inverse_stft / np.mean(np.abs(inverse_stft))
signal_scaled = signal / np.mean(np.abs(signal))
self.assertLess(np.std(inverse_stft_scaled - signal_scaled), threshold)
# Check that the inverse with correction and original signal are close.
self.assertLess(np.std(inverse_stft_corrected - signal),
corrected_threshold)
def _compare(self, signal, frame_length, frame_step, fft_length):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)) as sess:
actual_stft = spectral_ops.stft(
signal, frame_length, frame_step, fft_length, pad_end=False)
actual_inverse_stft = spectral_ops.inverse_stft(
actual_stft, frame_length, frame_step, fft_length)
actual_stft, actual_inverse_stft = sess.run(
[actual_stft, actual_inverse_stft])
expected_stft = SpectralOpsTest._np_stft(
signal, fft_length, frame_step, frame_length)
self.assertAllClose(expected_stft, actual_stft, 1e-4, 1e-4)
expected_inverse_stft = SpectralOpsTest._np_inverse_stft(
expected_stft, fft_length, frame_step, frame_length)
self.assertAllClose(
expected_inverse_stft, actual_inverse_stft, 1e-4, 1e-4)
def _compare(self, signal, frame_length, frame_step, fft_length):
with spectral_ops_test_util.fft_kernel_label_map(), (self.test_session(
use_gpu=True)) as sess:
actual_stft = spectral_ops.stft(signal,
frame_length,
frame_step,
fft_length,
pad_end=False)
actual_inverse_stft = spectral_ops.inverse_stft(
actual_stft, frame_length, frame_step, fft_length)
actual_stft, actual_inverse_stft = sess.run(
[actual_stft, actual_inverse_stft])
expected_stft = SpectralOpsTest._np_stft(signal, fft_length,
frame_step, frame_length)
self.assertAllClose(expected_stft, actual_stft, 1e-4, 1e-4)
expected_inverse_stft = SpectralOpsTest._np_inverse_stft(
expected_stft, fft_length, frame_step, frame_length)
self.assertAllClose(expected_inverse_stft, actual_inverse_stft,
1e-4, 1e-4)
def _compare_round_trip(self, signal, frame_length, frame_step, fft_length):
with spectral_ops_test_util.fft_kernel_label_map(), (
self.test_session(use_gpu=True)) as sess:
stft = spectral_ops.stft(signal, frame_length, frame_step, fft_length,
pad_end=False)
inverse_stft = spectral_ops.inverse_stft(stft, frame_length, frame_step,
fft_length)
signal, inverse_stft = sess.run([signal, inverse_stft])
# Since the shapes can differ due to padding, pad both signals to the max
# of their lengths.
max_length = max(signal.shape[0], inverse_stft.shape[0])
signal = np.pad(signal, (0, max_length - signal.shape[0]), "constant")
inverse_stft = np.pad(inverse_stft,
(0, max_length - inverse_stft.shape[0]), "constant")
# Ignore the frame_length samples at either edge.
start = frame_length
end = signal.shape[0] - frame_length
ratio = signal[start:end] / inverse_stft[start:end]
# Check that the inverse and original signal are equal up to a constant
# factor.
self.assertLess(np.var(ratio), 2e-5)
