[MRG+1] Added _fit_svd_solver variable to PCA #11225
Conversation
_fit_method will be set as soon as fit(X) method is called. Depending upon the value passes for svd_solver and input array X, value of _fit_method will be set.
sklearn/decomposition/pca.py
Outdated
| @@ -388,7 +388,7 @@ def _fit(self, X): | |||
| else: | |||
| n_components = self.n_components | |||
|
|
|||
| # Handle svd_solver | |||
| # Handle svd_solver and _fit_method | |||
| svd_solver = self.svd_solver | |||
There was a problem hiding this comment.
We can replace variable svd_solver with the new self._fit_method
There was a problem hiding this comment.
Everywhere? IMO svd_solver is more concise/readable..
There was a problem hiding this comment.
svd_solver is just being a local variable in _fit() method. If we replace it with a class variable (say self._fit_svd_solver for now), it will serve both the purposes, doing svd_solver's current task and expose the solver used.
Moreover, NearestNeighbours also handle algorithm='auto' in this manner.
There was a problem hiding this comment.
Yes, but you would have to do 4-5 replacements of svd_solver with self._fit_svd_solver to save one assignment, self._fit_svd_solver = svd_solver at the end (current solution). In any case that's not too important, either way would be fine, up to you.
|
Generally we need tests for a PR related to source code, but since |
|
I'm not sure that _fit_method is a great name... I'd rather have "solver in
there"
|
|
@jnothman Will it be okay if we overwrite value of |
No that wont work, because, when running, pca = PCA(svd_solver='auto')
pca.fit(X1)
pca.fit(X2)you wouldn't want the solver in the second fit to depend on the choice in the first one.
+1 maybe
Yes, though a test for two repeated fits (see above) behave consistently for the solver wouldn't hurt. Unless there is already a test like that. |
I'm OK with it and actually don't care much about the name (since it's private). Maybe we can also consider |
Replacing svd_solver with self._fit_svd_solver let us keep track of svd_solver method even if it is passed as 'auto'
sklearn/decomposition/pca.py
Outdated
| @@ -388,26 +388,26 @@ def _fit(self, X): | |||
| else: | |||
| n_components = self.n_components | |||
|
|
|||
| # Handle svd_solver | |||
There was a problem hiding this comment.
Not a serious problem but I think you might leave the comment as it is. (Handle _fit_svd_solver seems awkward from my side.)
sklearn/decomposition/pca.py
Outdated
| # Handle svd_solver | ||
| svd_solver = self.svd_solver | ||
| if svd_solver == 'auto': | ||
| # Handle _svd_solver |
There was a problem hiding this comment.
_svd_solver -> svd_solver :)
There was a problem hiding this comment.
@qinhanmin2014 Fixed typo.
|
@qinhanmin2014 @rth Would you guys want me to change anything further? |
There was a problem hiding this comment.
LGTM, merging. Thank you @njkevlani !
(I would have preferred to the keep the local svd_solver variable as less verbose, but since there is no consensus on that, this is fine.)
Fixes #11223 Expose solver used with PCA(svd_solver='auto')
_fit_method will be set as soon as fit(X) method is called. Depending upon the value passes for svd_solver and input array X, value of _fit_method will be set.