Initial implementation of GridFactory #21784
Conversation
|
I think that I did not comment on the issue but I really like what this factory remove as cognitive overhead on the user. The I would be happy to review and give feedback whenever you want. Just ping me directly. |
|
Wouldn't it make sense to add methods directly to the estimators them-selves? >>> pipeline = make_pipeline(
... PCA().with_search_space(n_components=[2, 5, 10, 20, 50, 100]),
... LogisticRegression().with_search_space(C=np.logspace(-6, 6, num=13)),
... )
>>> GridSearchCV(pipeline, cv=5).fit(X, y).best_params_
{"pca__n_components"=100, "logisiticregression__C"=1e-3}EDIT: One could even special-case pipelines to make it possible to accept tuple of estimators to express a disjunctions in the search space: >>> pipeline = make_pipeline(
... (
... PCA().with_search_space(n_components=[2, 5, 10, 20, 50, 100]),
... NMF().with_search_space(n_components=[2, 5, 10, 20, 50, 100]),
... ),
... LogisticRegression().with_search_space(C=np.logspace(-6, 6, num=13)),
... )
>>> GridSearchCV(pipeline, cv=5).fit(X, y).best_params_
{"nmf__n_components": 100, "logisiticregression__C": 1e-3}If you call This is a bit implicit, but given the gain in readability that might be worth it. |
|
BTW, another source of inspiration to tackle this problem: |
|
Thanks for those references, I'll enjoy reading them. I don't mind setting directly on the estimators, either with a method on the estimators or with an external function that just assumes it can write to a certain attribute of the estimator. Do we think this would be better? |
|
Setting an attribute could be a way to go toward an API where estimators could define their own default hyper parameters space in the future?
|
|
I agree it would probably be more intuitive to use. |
|
Neuraxle has some nice solutions, recognising some of the longstanding questionable design choices in Scikit-learn. One of our troubles is again going to be |
|
I agree. Let's put this API design on the agenda of the next dev meeting if you plan to attend @jnothman. |
|
Setting attributes not in For this PR, we can use the |
They support lists of search spaces, which is the same as GridSearchCV, right? |
I think what I find slightly counter-intuitive about that is that |
Reference Issues/PRs
Fixes #19045
What does this implement/fix? Explain your changes.
This implements a GridFactory with functionality comparable to searchgrid to allow easier specification of parameter search spaces, especially as a composite estimator is built.
Any other comments?
A complete version of this will:
cv_results_in a__-free wayI note that supporting a factory to be passed to
GridSearchCVinstead of a parameter grid doesn't save the user much code. Maybe we don't need to support it?