Using Several Parameters with GridSearchCV

[retkowski]

When using (for example) the following transformators

• CountVectorizer
• TruncatedSVD
• SelectKBest

with GridSearchCV it happens that it chooses a number for n_features for CountVectorizer that is less than n_components for TruncatedSVD or k for SelectKBest.

This leads to an error: ValueError: n_components must be < n_features

For SelectKBest I found a temporary solution:

class SelectAtMostKBest(SelectKBest):
    def _check_params(self, X, y):
        if not (self.k == "all" or 0 <= self.k <= X.shape[1]):
            self.k = "all"

But there is no equivalent for TruncatedSVD.
Is this behaviour intended? If yes, what can I do about this?

[jnothman]

Yes, this is a limitation of our API currently. One option is to use a parameter grid that only allows valid combinations, by using a list of parameter dicts wherein each setting is valid. A more robust forward-thinking solution might consider n_components and k optionally being functions of X and y.

[retkowski]

@jnothman Thank you, can you please elaborate a little bit more on the function-based approach? I am also considering to use the RandomizedSearchCV, but will probably face the same problem there, right?

[jnothman]

With randomized search you can't encode constraints between multiple parameters, but you can use error_score=0. The function-based approach would require someone changing the n_components interface to allow it to be a function. It's not something we've generally done, so it's probably on the scale of requiring an Enhancement Proposal to suggest that this approach be used across the board. You could potentially do something hacky like inherit from TruncatedSVD and overwrite fit(X, y), such that it changes n_components depending on X.shape.

…

On 31 January 2017 at 10:04, Fabian Retkowski ***@***.***> wrote: @jnothman <https://github.com/jnothman> Thank you, can you please elaborate a little bit more on the function-based approach? I am also considering to use the RandomizedSearchCV, but will probably face the same problem there, right? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#8243 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6zL8M94v7cetnyu1-urZvHQCvUctks5rXmxkgaJpZM4Lw5tC> .

[amueller]

How about using SelectPercentile instead? Not sure if we allow covered variance in TruncatedSVD but it would be nice to have an option that is relative to the input feature size, not absolute.