ENH CountVectorizer: sort features after pruning by frequency

[smola]

In CountVectorizer, sort features after pruning by frequency.
When using min_df or max_df with large vocabularies, this can
be a significant speed up.

Here's a simple benchmark:

from sklearn.datasets import fetch_20newsgroups

dataset = fetch_20newsgroups(subset="all")

from time import time
from sklearn.feature_extraction.text import CountVectorizer

n_doc = 2000

for min_df in (1, 2, 3, 5, 8):
    for min_ngram_range in (1,):
        for max_ngram_range in (min_ngram_range + 5,):
            t0 = time()
            cv = CountVectorizer(
                input="content",
                analyzer="char",
                lowercase=False,
                ngram_range=(min_ngram_range, max_ngram_range),
                min_df=min_df,
            )
            cv.fit(dataset.data[:n_doc])
            print(
                "time=%f min_df=%d ngram_range=(%d, %d)"
                % (time() - t0, min_df, min_ngram_range, max_ngram_range)

and example results before the change:

time=20.674258 min_df=1 ngram_range=(1, 6)
time=21.181467 min_df=2 ngram_range=(1, 6)
time=21.105731 min_df=3 ngram_range=(1, 6)
time=21.049018 min_df=5 ngram_range=(1, 6)
time=21.095739 min_df=8 ngram_range=(1, 6)

and after the change:

time=20.434718 min_df=1 ngram_range=(1, 6)
time=17.153750 min_df=2 ngram_range=(1, 6)
time=16.405343 min_df=3 ngram_range=(1, 6)
time=16.040197 min_df=5 ngram_range=(1, 6)
time=15.901990 min_df=8 ngram_range=(1, 6)

[jnothman]

Great idea! Not sure how we missed this when we cleaned this code up a couple of years ago...

Please add an Efficiency entry to the change log at doc/whats_new/v0.23.rst. Like the other entries there, please reference this pull request with :pr: and credit yourself (and other contributors if applicable) with :user:

[smola]

@jnothman Thank you for the review. I have rebased and added the changelog entry.

[jnothman]

No need to rebase. Please just append commits.

Awaiting another review.

[rth]

Thanks @smola , LGTM.