python - astropy.convolution.convolve returns nan values

Question

I am trying to use the astropy module to smooth my data. As a 1D example I have tried the following code:

import numpy as np
from astropy import convolution as conv
var1=np.arange(10)
kernel=np.asarray([-1,1,0])
conv.convolve(var1,kernel)

This returns in my case: array([ nan, nan, nan, nan, nan, nan, nan, nan, nan, nan])

I assumed that it should return array([0,1,1,1,1,1,1,1,1,1]). I have tried using for var1 datatypes uint8, int8 and float32. I have also tried using convolve_fft, with the same result.

I am familiar with the convolution filter of scipy but I want to use astropy for when NaN values are actually present in my data, so as to smooth them over.

Answer 1

您的内核总和为零，因此astropy会发出警告

RuntimeWarning: invalid value encountered in true_divide kernel_internal /= kernel_sum

最终导致所有nans.

Answer 2

正如已经指出的那样，这是因为您的内核的总和为零：

>>> import numpy as np
>>> np.sum(kernel)  # kernel = np.asarray([-1,1,0])
0

kernel如果不规范化但astropy.convolution.convolve总是规范化内核以插入NaN数组中的值（因为 astropy 1.3 也被屏蔽）并将结果再次乘以原始内核的总和，这本身不会成为问题（除非你明确使用normalize_kernel=True）。

---

即使您已经说过您熟悉它，您也可以始终使用scipy.ndimage.convolve：

>>> from scipy.ndimage import convolve

>>> convolve(var1, kernel[::-1])  # var1=np.arange(10)
array([0, 1, 1, 1, 1, 1, 1, 1, 1, 1])

并自己进行插值和卷积以及NaN位置！（不是很好，但是对于零和、不对称和混合签名的内核，你不能正确地进行卷积）。

[::-1]但是，在使用scipy时，您需要反转（因此是）内核。

在您的情况下，您也可以简单地使用numpy.diff（删除前导零）：

>>> np.diff(var1)
array([1, 1, 1, 1, 1, 1, 1, 1, 1])