Thanks for the pull-request !
Could you add a small test in sklearn/feature_extraction/tests/test_dict_vectorizer.py ?

Could you add a small test in sklearn/feature_extraction/tests/test_dict_vectorizer.py ?

The problem is that this only happens for sparse arrays that have more than 2**31 elements which corresponds to 10s GB RAM. I think it's OK to merge it without unit test given that's idential to what we did earlier for CountVectorizer in #9147

@norvan Could you please run a small performance benchmark before and after this PR to make sure performance is not degraded? It shouldn't be, but better to double check.

For instance running evaluation code in jupyter with %timeit DictVectorizer should be enough. Thanks!

Ran the following code on both stable sklearn and after my changes.
Duration was:

• 2.79 s ± 16.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) before the change
• 2.83 s ± 25.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) after the change

Conclusion is that the change does not impair performance in a statistically meaningful way considering the tolerance.

from sklearn.feature_extraction import DictVectorizer
from sklearn.feature_selection import GenericUnivariateSelect, chi2
from numpy.random import random

n_samples = 1000000

data = [{"alpha": i, "beta": int(random()*100)} for i in range(n_samples)]

vectorizer = DictVectorizer(sparse=True, dtype=np.uint16).fit(data)

%timeit vectorized_data = vectorizer.transform(data)