Lars Grüter

Superseded by https://github.com/scikit-image/scikit-image/pull/7350.

Hey, sorry for our delayed response on this. I'll have a look.

I am noticing that for the old and new implementation the number seems to change a bit. E.g. for the example below the residuals change on a scale

With ```python import numpy as np import skimage as ski rng = np.random.default_rng(42) src = rng.random((100, 2)) dst = rng.random((100, 2)) tform = ski.transform.PiecewiseAffineTransform() tform.estimate(src, dst) x = rng.random((2000, 2))...

Thanks @svepe! I tried to address the test failure in https://github.com/scikit-image/scikit-image/pull/6963/commits/949cf8a3b12e2d729ed658b90f43986ecdff2cdb.

@svepe, I think this looks good. However, I am hesitant to approve this because this performance optimization does change our results, if only slightly. We have a policy to to...

Point in favor of merging this is that the snippet below seems to pass always. We use `assert_allclose` elsewhere to check our output with a relative tolerance of 1e-7. So...

I think we have to wait a bit until this bug https://github.com/pydata/pydata-sphinx-theme/pull/1450 is resolved...

> We're going to try to do this in SciPy once we update our theme version (hopefully imminent). Maybe I'm dense, what will you try? Manual backporting or rebuilding the...

cc'ing @jni and @stefanv, since both of were involved in https://github.com/scikit-image/scikit-image/issues/7023. If you have a bit of time left, this should be a simple review. See https://github.com/scikit-image/scikit-image/issues/7023#issuecomment-1594743246 for an explanation...