Use scipy.signal.correlate

[ahwillia]

A small suggestion I have is to simplify the fft calls where possible. For example you should be able to replace the following line:

muse-python/muse/util.py

Line 394 in d4e5ab1

xcorr_raw_this = np.fft.fftshift(np.real(np.fft.ifft(Xcorr_raw_this_padded)))

With a simpler call to scipy.signal.correlate. Note that you can use method='fft' to force this function to compute the cross-correlation in the frequency domain. By default, method='auto' will take a good guess for which method is faster.

I'm not sure you'll see big performance gains, but its possible.

[amanchoudhri]

Good call on trying to limit fft. I used a python standard library module called cProfile to get an in-depth breakdown of the most costly functions called, and it turns out that calls to ifft account for almost 90% of the running time! I'll try again with the function you linked and let you know what happens.

[ahwillia]

I suspect that it is probably inescapable. It is probably a good sign that fft / ifft are taking up that much time, as they are heavily optimized. Nonetheless, using scipy.signal.correlate is simpler and less error-prone.