Switch from GPfit to GauPro for Gaussian Processes #1104
Conversation
| @@ -1,4 +1,17 @@ | |||
| library(tidymodels) | |||
| library(broom) | |||
There was a problem hiding this comment.
Note: we were getting warnings with the old tests that:
namespace 'workflowsets' is not available and has been replaced
Not loading tidymodels solves this and it comes up in the tests for #1007 1007
| "There are {cli::qty(diff)}{?as many/more} tuning parameters | ||
| {cli::qty(diff)}{?as/than} there are initial points. |
There was a problem hiding this comment.
Why are these lines being deleted?
if that is intentional then we also don't need cli::pluralize()
There was a problem hiding this comment.
We have this warning now. which I find more informative. I don't think that we need both.
msg <- cli::format_inline(
"The Gaussian process model is being fit using {num_pred} feature{?s} but
only has {num_cand} data point{?s} to do so. This may cause errors or a
poor model fit."
)| dplyr::filter(.metric == metrics) |> | ||
| dplyr::filter(!is.na(mean)) | ||
|
|
||
| # TODO a lot of slice_min/slice_max can be used now |
There was a problem hiding this comment.
Should add as an issue if you are not planning on dealing with it in this PR.
R/gp_helpers.R
Outdated
| } | ||
|
|
||
| withr::with_seed( | ||
| 114, |
There was a problem hiding this comment.
why are we hardcoding this specific seed?
There was a problem hiding this comment.
That is a leftover from development. I've taken it out.
The GPfit maintainer contacted me a few months ago about the potential for that package to fall off CRAN. This change helps from a sustainability point of view, but also adds some new features.
I've tested this extensively on the analyses in aml4td and TMwR and get very similar results.