Multi-dimension 1D FFT

1

I am trying to convolve several 1D signals via FFT convolution. Does Pytorch offer any ways to avoid a for loop as below to perform a multi-dimension 1D FFT / iFFT, i.e. I would like to have a batch-wise 1D FFT?

import torch

# 1D convolution (mode = full)
def fftconv1d(s1, s2):
    # extract shape
    nT = len(s1)
    # signal length
    L = 2 * nT - 1
    # compute convolution in fourier space
    sp1 = torch.fft.fft(s1, L)
    sp2 = torch.fft.fft(s2, L)
    conv = torch.real(torch.fft.ifft(sp1 * sp2))
    return conv

A = torch.randn(20, 25)
B = torch.randn(20, 25)
conv = torch.randn(20, 25*2 -1)

for it in range(20):
    conv[it, :] = fftconv1d(A[it, :], B[it, :])
2

Your method should work if you fix the nT calculation and allows you to pass the inputs directly:

# 1D convolution (mode = full)
def fftconv1d(s1, s2):
    # extract shape
    nT = s1.size(-1)#len(s1)
    # signal length
    L = 2 * nT - 1
    # compute convolution in fourier space
    sp1 = torch.fft.fft(s1, L)
    sp2 = torch.fft.fft(s2, L)
    conv = torch.real(torch.fft.ifft(sp1 * sp2))
    return conv

A = torch.randn(20, 25)
B = torch.randn(20, 25)
conv = torch.randn(20, 25*2 -1)

for it in range(20):
    conv[it, :] = fftconv1d(A[it, :], B[it, :])

out = fftconv1d(A, B)
print((conv - out).abs().max())
# tensor(6.2585e-07)
3

I didnt notice that. Great thanks!!!