def test_correlations_narrow_v1_vs_v5(self):
inverse_power = wpe.get_power_inverse(self.Y)
R_desired, r_desired = wpe.get_correlations_narrow(
self.Y, inverse_power, self.K, self.delay)
R_actual, r_actual = wpe.get_correlations_narrow_v5(
self.Y, inverse_power, self.K, self.delay)
tc.assert_allclose(R_actual, R_desired)
tc.assert_allclose(r_actual, r_desired)
def test_inverse_power(self):
np_inv_power = wpe.get_power_inverse(self.Y)
with self.test_session() as sess:
tf_signal = tf.placeholder(tf.complex128, shape=[None, None])
tf_res = tf_wpe.get_power_inverse(tf_signal)
tf_inv_power = sess.run(tf_res, {tf_signal: self.Y})
np.testing.assert_allclose(np_inv_power, tf_inv_power)
def test_correlations_v1_vs_v2_toy_example(self):
K = 3
delay = 1
Y = np.asarray([[11, 12, 13, 14], [41, 22, 23, 24]], dtype=np.float32)
inverse_power = wpe.get_power_inverse(Y)
R_desired, r_desired = wpe.get_correlations(Y, inverse_power, K, delay)
R_actual, r_actual = wpe.get_correlations_v2(Y, inverse_power, K,
delay)
tc.assert_allclose(R_actual, R_desired)
tc.assert_allclose(r_actual, r_desired)
def test_filter_matrix_conj_v1_vs_v5(self):
inverse_power = wpe.get_power_inverse(self.Y)
correlation_matrix, correlation_vector = wpe.get_correlations(
self.Y, inverse_power, self.K, self.delay)
desired = wpe.get_filter_matrix_conj(correlation_matrix,
correlation_vector, self.K,
self.D)
actual = wpe.get_filter_matrix_conj_v5(self.Y, inverse_power, self.K,
self.delay)
tc.assert_allclose(actual, desired, atol=1e-10)
def test_correlations(self):
np_inv_power = wpe.get_power_inverse(self.Y)
np_corr = wpe.get_correlations_narrow(self.Y, np_inv_power, self.K,
self.delay)
with tf.Graph().as_default(), tf.Session() as sess:
tf_signal = tf.placeholder(tf.complex128, shape=[None, None])
tf_inverse_power = tf_wpe.get_power_inverse(tf_signal)
tf_res = tf_wpe.get_correlations_for_single_frequency(
tf_signal, tf_inverse_power, self.K, self.delay)
tf_corr = sess.run(tf_res, {tf_signal: self.Y})
np.testing.assert_allclose(np_corr[0], tf_corr[0])
np.testing.assert_allclose(np_corr[1], tf_corr[1])
def test_correlations_narrow_v1_vs_v6(self):
inverse_power = wpe.get_power_inverse(self.Y)
R_desired, r_desired = wpe.get_correlations_narrow(
self.Y, inverse_power, self.K, self.delay)
s = [Ellipsis, slice(self.delay + self.K - 1, None)]
Y_tilde = wpe.build_y_tilde(self.Y, self.K, self.delay)
R_actual, r_actual = wpe.get_correlations_v6(self.Y[s], Y_tilde[s],
inverse_power[s])
tc.assert_allclose(R_actual.conj(), R_desired)
tc.assert_allclose(r_actual.conj(),
np.swapaxes(r_desired, 1,
2).reshape(-1, r_desired.shape[-1]),
rtol=1e-5,
atol=1e-5)
def test_filter_matrix(self):
np_inv_power = wpe.get_power_inverse(self.Y)
np_filter_matrix = wpe.get_filter_matrix_conj_v5(
self.Y, np_inv_power, self.K, self.delay)
with tf.Graph().as_default(), tf.Session() as sess:
tf_signal = tf.placeholder(tf.complex128, shape=[None, None])
tf_inverse_power = tf_wpe.get_power_inverse(tf_signal)
tf_matrix, tf_vector = tf_wpe.get_correlations_for_single_frequency(
tf_signal, tf_inverse_power, self.K, self.delay)
tf_filter = tf_wpe.get_filter_matrix_conj(tf_signal, tf_matrix,
tf_vector, self.K,
self.delay)
tf_filter_matrix, tf_inv_power_2 = sess.run(
[tf_filter, tf_inverse_power], {tf_signal: self.Y})
np.testing.assert_allclose(np_inv_power, tf_inv_power_2)
np.testing.assert_allclose(np_filter_matrix, tf_filter_matrix)
def test_filter_matrix_conj_v1_vs_v7(self):
inverse_power = wpe.get_power_inverse(self.Y)
correlation_matrix, correlation_vector = wpe.get_correlations(
self.Y, inverse_power, self.K, self.delay)
desired = wpe.get_filter_matrix_conj(correlation_matrix,
correlation_vector, self.K,
self.D)
s = [Ellipsis, slice(self.delay + self.K - 1, None)]
Y_tilde = wpe.build_y_tilde(self.Y, self.K, self.delay)
actual = wpe.get_filter_matrix_v7(
self.Y,
Y_tilde=Y_tilde,
inverse_power=inverse_power,
)
tc.assert_allclose(actual.conj(),
np.swapaxes(desired, 1,
2).reshape(-1, desired.shape[-1]),
atol=1e-10)
def test_filter_operation(self):
np_inv_power = wpe.get_power_inverse(self.Y)
np_filter_matrix = wpe.get_filter_matrix_conj_v5(
self.Y, np_inv_power, self.K, self.delay)
np_filter_op = wpe.perform_filter_operation_v4(self.Y,
np_filter_matrix,
self.K, self.delay)
with tf.Graph().as_default(), tf.Session() as sess:
tf_signal = tf.placeholder(tf.complex128, shape=[None, None])
tf_inverse_power = tf_wpe.get_power_inverse(tf_signal)
tf_matrix, tf_vector = tf_wpe.get_correlations_narrow(
tf_signal, tf_inverse_power, self.K, self.delay)
tf_filter = tf_wpe.get_filter_matrix_conj(tf_signal,
tf_inverse_power,
tf_matrix, tf_vector,
self.K, self.delay)
tf_filter_op = tf_wpe.perform_filter_operation(
tf_signal, tf_filter, self.K, self.delay)
tf_filter_op = sess.run(tf_filter_op, {tf_signal: self.Y})
np.testing.assert_allclose(np_filter_op, tf_filter_op)
