[英]Need a circular FFT convolution in Python
I need a faster analog of 我需要更快的模拟
scipy.signal.convolve2d(data, filter, boundary="wrap", mode="same")
Cannot you advice me how to replace it? 您不能建议我如何更换它吗?
PS scipy.signal.fftconvolve
is fast enough, but it does not have boundary
option and I cannot make it work in circular convolution mode. PS scipy.signal.fftconvolve
足够快,但是没有boundary
选项,因此我无法使其在循环卷积模式下工作。
If you compute the following: 如果您计算以下内容:
from scipy.fftpack import fft2, ifft2
f2 = ifft2(fft2(data, shape=data.shape) * fft2(filter, shape=data.shape)).real
then f2
contains the same values as convolve2d(data, filt, boundary='wrap', mode='same')
, but the values are shifted ("rolled", in numpy terminology) in each axis. 然后f2
包含与convolve2d(data, filt, boundary='wrap', mode='same')
的值,但是这些值在每个轴上移动(以“ numpy”术语表示为“ rolled”)。 (This is an application of the convolution theorem .) (这是卷积定理的一个应用。)
Here's a short function that rolls the result to the give same result as the convolve2d
function call: 这是一个简短的函数,可将结果转换为与convolve2d
函数调用相同的结果:
def fftconvolve2d(x, y):
# This assumes y is "smaller" than x.
f2 = ifft2(fft2(x, shape=x.shape) * fft2(y, shape=x.shape)).real
f2 = np.roll(f2, (-((y.shape[0] - 1)//2), -((y.shape[1] - 1)//2)), axis=(0, 1))
return f2
For example, 例如,
In [91]: data = np.random.rand(256, 256)
In [92]: filt = np.random.rand(16, 16)
In [93]: c2d = convolve2d(data, filt, boundary='wrap', mode='same')
In [94]: f2 = fftconvolve2d(data, filt)
Verify that the results are the same: 验证结果是否相同:
In [95]: np.allclose(c2d, f2)
Out[95]: True
Check the performance: 检查性能:
In [96]: %timeit c2d = convolve2d(data, filt, boundary='wrap', mode='same')
44.9 ms ± 77.3 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [97]: %timeit f2 = fftconvolve2d(data, filt)
5.23 ms ± 11.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
The FFT version is much faster (but note that I chose the dimensions of data
to be a power of 2). FFT版本要快得多(但请注意,我选择的data
尺寸为2的幂)。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.