A package that aims at simplifying the usage of FFT in Keras3.
Keras3 does not have a complex dtype. This means, that the ops.fft call is a bit cumbersome, as it expects a tuple of real- and imaginary part in float32.
The basic FFT part of this package acts as an inplace option for ops.fft, which handles the input automatically.
It accepts
- a tuple of real and imaginary part,
fft((x_real, x_imag)), or - a single float KerasTensor, which is then interpreted as the real part,
fft(x).
The latter option automatically initializes a zero-Tensor with the same shape and dtype as x.
Additionally, the module keras_fft.derivative contains code for the differentiation in Fourier space,
which is an elegant way to get the nth derivative of a signal.
Keras backends JAX and Tensorflow are currently supported.