MNT Use const memory views in DistanceMetric subclasses #19883
Conversation
|
Const fused types has not been released yet. But in this case, |
const with fused types memory views for cluster.DistanceMetricsconst memory views for cluster.DistanceMetrics
ogrisel
left a comment
There was a problem hiding this comment.
+1 for fixing #19875 explicitly as part of this PR with a dedicated non-regression test + an entry in the changelog.
Since the DistanceMetric class is part of the documented public API, we should also mention the fix for this class in the changelog.
Co-authored-by: Olivier Grisel <[email protected]>
ogrisel
left a comment
There was a problem hiding this comment.
Some minor comments but otherwise LGTM. Is this PR style a DRAFT? Do you plan other changes? I think it's already a net improvement.
Co-authored-by: Olivier Grisel <[email protected]>
Co-authored-by: Olivier Grisel <[email protected]>
Co-authored-by: Olivier Grisel <[email protected]>
Fix finting, again.
I was inspecting if changes made to |
const memory views for cluster.DistanceMetricsconst memory views for neighbors.DistanceMetric subclasses
thomasjpfan
left a comment
There was a problem hiding this comment.
Thank you for the PR @jjerphan !
Co-authored-by: Thomas J. Fan <[email protected]>
const memory views for neighbors.DistanceMetric subclasses…#19883) Co-authored-by: Olivier Grisel <[email protected]> Co-authored-by: Thomas J. Fan <[email protected]>
Reference Issues/PRs
Fixes #19875.
See also #10624.
What does this implement/fix? Explain your changes.
Distances computations (
cluster.DistanceMetrics) implemented in Cython do not support readonly datasets, which are obtained when using mem-mapping for instance. This makes some interfaces relying on those implementation fail (see #19875).This PR makes the Cython implementations use
constmemoryviews to be able computing those distances on memmapped datasets.Any other comments?
5b61a29 fixes #19875, but maybe we should have the downstream issues be solved independently from this PR.
Also
scipy.spatial.distance.{cdist,pdist}on memmapped datasets with'mahalanobis'as a metric fail.I am inspecting why; in the meantime 716b330 skips the associated tests.