Benchmarks

This notebook contains benchmarks against the pynfft package, which is a Python wrapper of Fortran code



In [1]:

    
import numpy as np



In [2]:

    
# Import from main directory
import os, sys
sys.path.append(os.path.abspath('..'))

import nfft
nfft.__version__









    Out[2]:





'0.1'



In [3]:

    
import pynfft
pynfft.__version__









    Out[3]:





'1.3.2'

Benchmarking the Forward Transform

Define some test data:



In [4]:

    
def make_forward_data(M, N):
    x = -0.5 + np.random.rand(M)
    f_hat = np.random.randn(N) + 1j * np.random.randn(N)
    return x, f_hat

Define a utility function around pynfft:



In [5]:

    
def pynfft_forward(x, f_hat):
    M = len(x)
    N = len(f_hat)
    plan = pynfft.nfft.NFFT(N, M)
    plan.x = x
    plan.precompute()
    plan.f_hat = f_hat
    # Need copy because of bug in pynfft 1.x
    # See https://github.com/ghisvail/pyNFFT/issues/57
    return plan.trafo().copy()



In [6]:

    
x, f_hat = make_forward_data(1000, 100000)



In [7]:

    
out1 = nfft.nfft(x, f_hat)
out2 = pynfft_forward(x, f_hat)
np.allclose(out1, out2)









    Out[7]:





True



In [8]:

    
%timeit nfft.nfft(x, f_hat)
%timeit pynfft_forward(x, f_hat)









    



10 loops, best of 3: 22.4 ms per loop
10 loops, best of 3: 43.1 ms per loop

Benchmarking the Adjoint Transform

Define some test data:



In [9]:

    
def make_adjoint_data(M):
    x = -0.5 + np.random.rand(M)
    f = np.random.randn(M) + 1j * np.random.randn(M)
    return x, f

Define a utility function around pynfft:



In [10]:

    
def pynfft_adjoint(x, f, N):
    M = len(x)
    plan = pynfft.nfft.NFFT(N, M)
    plan.x = x
    plan.precompute()
    plan.f = f
    # Need copy because of bug in pynfft 1.x
    # See https://github.com/ghisvail/pyNFFT/issues/57
    return plan.adjoint().copy()



In [11]:

    
x, f = make_adjoint_data(1000)
N = 100000



In [12]:

    
out1 = nfft.nfft_adjoint(x, f, N)
out2 = pynfft_adjoint(x, f, N)

np.allclose(out1, out2)









    Out[12]:





True



In [13]:

    
%timeit nfft.nfft_adjoint(x, f, N, sigma=3)
%timeit pynfft_adjoint(x, f, N)









    



10 loops, best of 3: 24.2 ms per loop
10 loops, best of 3: 42.8 ms per loop