While you are at it, could you please edit the comments to expand the CI acronym from "because of CI’s time constraints" to "because of resource usage constraints on our Continuous Integration infrastructure that is used to build this documentation on a regular basis."

The comment in the code says that this doesn't converge, does it still not converge?

There's also a mistake in the comment, there's an extra "are" there, you can fix that too.

Also, please mention the absolute times before and after.

@ogrisel Edited your improved explanation into the docstring

@adrinjalali Yes, rest should be good, will delete the additional "are" and also paste in some new run times

@adrinjalali @ogrisel I deleted the additional "are" and added @ogrisel`s comment also in that line since it was the same as in the docstring, so I figured to put the new explanation in both lines (docstring and single line comment)

I think we can get a better speedup here if we use a much smaller part of the data as training set. Could you please use train_test_split, and use a small part of the data as training, while somewhat keeping the output the same?

@adrinjalali Okay, I'll get on it, soon :)

@adrinjalali We get the following results:

With this code

X_train, X_test = X[:60000], X[60000:]
y_train, y_test = y[:60000], y[60000:]

we get this result:

and with this code

X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0, test_size=0.7)

we get this result:

So, you were right. train_test_split is great. I tried going with less than 30 % train data but it doesn't get much faster, only the validation accuracy gets slightly worse (at least different from the original behavior)...

Looking at the changes since @ogrisel 's approval, I think we can merge, thanks @sveneschlbeck