Repository navigation
Issues with NaN treatment and normalisation in convolve_fft #8114
Description
convolve_fft(nan_treatment='fill') doesn't normalize when passed normalization function
When no NaN or inf are present in the input arrays, nan_treatment=('fill,'interpolate') should have identical behavior. The following test fails.
import numpy as np
from astropy.convolution import convolve, convolve_fft
from astropy.convolution.tests.test_convolve_fft import assert_floatclose
array = [1, 2, 3]
kernel = [3, 3, 3]
fill = convolve_fft(array, kernel, normalize_kernel=np.max, nan_treatment='fill')
interpolate = convolve_fft(array, kernel, normalize_kernel=np.max, nan_treatment='interpolate')
assert_floatclose(fill, interpolate)This is only an issue when passing a normalization function via normalize_kernel.
This is blocking #8101.
I am currently in the progress of fixing this issue and refactoring the scattered and implicit normalization logic.
Activity
jamienoss commentedon Nov 15, 2018
normalize_kernel=someFunc() and nan_treatment='interpolate'
Note: I'm not yet sure if this is a bug because I'm not yet 100% sure about how the interpolation works.
The following exception should be raised regardless of the state of
normalize_kernel, i.e. it should be pulled out and checked after the normalization block.[src]
... kernel_scale = kernel.sum() if np.abs(kernel_scale) < normalization_zero_tol: if nan_treatment == 'interpolate': raise ValueError('Cannot interpolate NaNs with an unnormalizable kernel') ...Furthermore, if
normalize_kernel=someFunc()andnan_treatment='interpolate'then the exception should not be conditional onkernel.sum()butsomeFunc(kernel). Also, any normalization used for the interpolation should usesomeFunc.I have assumed that it is and refactored it accordingly whilst working on this issue. If it turns out to really be a bug, I'll open an issue and link in the commit message, I'll remove otherwise.
jamienoss commentedon Nov 15, 2018
Working branch: master...jamienoss:fft-ker-norm-issue
[-]convolve_fft(nan_treatment='fill') doesn't normalize when passed normalization function[/-][+]Issues with NaN treatment and normalisation in convolve_fft[/+]