fourierWeight.py

import numpy as np
from keras import backend as K

FLENGTH = 500
FSTEP = 200


def fourierWeight(length, fraction=16.0):
    freq_length = length/fraction
    sample_index = np.array([range(0, length)], dtype=np.float32)
    freq_index = np.array([range(0, length)], dtype=np.float32)
    freq_index = freq_index/fraction

    w = np.matmul(np.transpose(sample_index), freq_index)
    w = w*(2*np.pi)/length  # phi = 2*PI*i*k/n
    sines = -np.sin(w)/length  # e(-j*phi) = -jsin(phi) + cos(phi)
    cosines = np.cos(w)/length

    # para-curve window  f(0) = 0; f(n-1) = 0;  f((n-1)/2) = 1
    window = np.array(range(0, length), dtype=np.float32)
    window = np.reshape(window, [length, 1])
    window = (window * 4.0 / (length - 1)) - np.multiply(window,
                                                         window) * 4.0 / ((length - 1) * (length - 1))
    window = window.astype(np.float32)
    sines = np.multiply(window, sines)
    cosines = np.multiply(window, cosines)
    sines = np.reshape(sines, [length, 1, 1, -1])
    cosines = np.reshape(cosines, [length, 1, 1, -1])
    return sines, cosines


def fourierLayer(x):
    sines, cosines = fourierWeight(FLENGTH, 4.0)
    s = K.conv2d(x, sines, (FSTEP, 1))
    c = K.conv2d(x, cosines, (FSTEP, 1))
    o = K.concatenate([s, c], axis=2)
    std = K.std(o)
    mean = K.mean(o)
    o = (o - mean)/std
    o = K.permute_dimensions(o, [0, 1, 3, 2])
    return o


def fourierLayerShape(x):
    if x[2] != 1 or x[3] != 1:
        raise ValueError(
            'input shape must be [sample, seq_length, 1, 1], but got ' + str(x))
    O = (x[0], None, FLENGTH, 2)
    return O


if __name__ == '__main__':
    fourierWeight(101)