def call(self, inputs):
mean = tf.math.reduce_mean(inputs, axis=self.axis, keepdims=True)
deviation_abs = tf.math.reduce_mean(tf.abs(inputs - mean),
axis=self.axis,
keepdims=True)
return (inputs - mean) / (deviation_abs + self.epsilon)
def _fft_magnitude(self, inputs):
"""Compute FFT and returns its magnitude."""
real_spectrum = tf.signal.rfft(inputs, [self.fft_size])
magnitude_spectrum = tf.abs(real_spectrum)
if self.magnitude_squared:
return tf.square(magnitude_spectrum)
else:
return magnitude_spectrum
def test_tf_fft_vs_rdft_direct(self):
signal_size = 64
# input signal
signal = np.random.rand(1, signal_size)
# build rfft model and run it
input_signal = tf.keras.Input(shape=(signal_size, ), batch_size=1)
spectrum = tf.signal.rfft(input_signal)
spectrum = tf.abs(spectrum)
model = tf.keras.Model(input_signal, spectrum)
model.summary()
spectrum_output = model.predict(signal)
# build rdft model and run it
input_signal = tf.keras.Input(shape=(signal_size, ), batch_size=1)
output = magnitude_rdft.MagnitudeRDFT(
magnitude_squared=False)(input_signal)
model = tf.keras.Model(input_signal, output)
model.summary()
rdft_output = model.predict(signal)
self.assertAllClose(rdft_output, spectrum_output)