-
In the Acoular package, we are using scikit-learn for some linear model solving. In our regression testing I came across a problem where I would appreciate some advise. When using This was not the case for prior versions of Acoular, when we used scikit-learn<1.2. Because of the change with normalize=True introduced with scikit-learn 1.2, I had to change our code, and despite some effort, I was not able with 1.2 to reproduce the results I got with <1.2. My solution was to change the code in Acoular , so that now the result is the same across all Python and sklearn versions on the same OS. I could make the test pass for all OS by increasing the test tolerance to say 1e-4 in this case, but before I do so I would rather like to know if there is a way to improve reproducibility itself in this case. Maybe there is some option or approach that I overlooked. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
I don't recall exactly the deprecation of We probably handle internally this case in a similar fashion by playing with the tolerance. |
Beta Was this translation helpful? Give feedback.
I don't recall exactly the deprecation of
normalize
but it might be possible that the version of theStandardScaler
will be slightly different because we don't use the same data to standardize at the end.We probably handle internally this case in a similar fashion by playing with the tolerance.