Hi @jrbourbeau
Thanks a lot for the patch. It looks good overall, but can you add a test?

(also, it might be more elegant to use the from __future__ import division solution.)

I would prefer the from __future__ import division route as well. @jnothman, do you think it's safe to assume that the internals of sklearn.decomposition.IncrementalPCA aren't something users will be doing a copy-paste with?

@NelleV, do have any suggestions for a test to add? I was thinking the original code that caught this problem in issue #9489 might be a good test. Something like

import numpy as np
from sklearn.decomposition import IncrementalPCA

def test_partial_fit_correct_answer():
    A = np.array([[1, 2, 4], [5, 3, 6]])
    B = np.array([[6, 7, 3], [5, 2, 1], [3, 5, 6]])
    C = np.array([[3, 2, 1]])

    ipca = IncrementalPCA(n_components=2)
    ipca.partial_fit(A)
    ipca.partial_fit(B)
    # Know answer is [[-1.48864923 -3.15618645]]
    np.testing.assert_equal(ipca.transform(C), [[-1.48864923 -3.15618645]])

But I'm not sure if hard-coding the correct answer is the way to go, or if one should just simply check for float division (since that's the real problem here) with something like

def test_float_division():
    assert 3/2 == 1.5

A test hard-coding the correct answer with a comment saying it's a non regression test for issue#9489 would be great

Here is a snippet that can be easily turned into a regression test:

import numpy as np
from sklearn.decomposition import IncrementalPCA

rng = np.random.RandomState(0)
A = rng.randn(5, 3) + 2
B = rng.randn(7, 3) + 5

pca = IncrementalPCA(n_components=2)
pca.partial_fit(A)
pca.n_samples_seen_ = float(pca.n_samples_seen_)
pca.partial_fit(B)
print(pca.singular_values_)

pca2 = IncrementalPCA(n_components=2)
pca2.partial_fit(A)
pca2.partial_fit(B)
print(pca2.singular_values_)

With scikit-learn 0.18.2 and Python 2.7, the output is:

[ 7.70269857  3.88561777]
[ 6.64122066  3.88184665]

Also you want to add an entry into doc/whats_new.rst about your fix. Do mention it is Python 2 only.

It looks like I added a bunch of other commits I didn't intend to in my most recent push. I was trying to update doc/whats_new.rst with an entry for this PR, but had to rebase on master to get the updates to doc/whats_new.rst that had been added since the feature branch for this PR was made.

I attempted the rebase following the instructions from contributing guide (https://docs.scipy.org/doc/numpy/dev/gitwash/development_workflow.html#rebasing-on-master), but must have accidentally included all the other commits as well.

I'm not exactly sure what to do here. Does anyone know of a way to fix this? Or should I open up a new PR? Thanks!!

I made a backup branch before I used git rebase. Would

# reset branch back to the saved point
git reset --hard backup_branch
# push this saved point to the remote branch for this PR
git push -f origin remote_feature_branch

work to revert this PR back to its state before I did the rebase?

Thanks so much @jnothman!

I want to add an entry to doc/whats_new.rst related to this PR. However, there have been changes made to the current master branch version of doc/whats_new.rst. Namely, a section for version 0.20 has been added. I would like to add an entry for this PR to the version 0.20 section without creating any merge conflicts with the current master branch.

I was wondering how I should go about this. One possible way would be to copy the new version 0.20 section from the current master branch into my feature branch, add an entry for this PR, and then commit it. Is this the preferred way, or is there a different way to go about this?

It worked! Thanks for helping me with that @jnothman

PR #9526 updated doc/whats_new.rst on master in preparation for the 0.19 release, leading to a merge conflict. But now all conflicts have been resolved. Unless there are any other comments, I think this PR is good to go.

To summarize, this PR:

1. Makes sure the partial_fit method for sklearn.decomposition.IncrementalPCA will use float
   division for Python 2 versions.
2. Adds non-regression test for issue IncrementalPCA.partial_fit doesn't use float division in python 2 #9489 to sklearn/decomposition/tests/test_incremental_pca.py.
3. Updates doc/whats_new.rst with an entry for this PR.

Thanks @jrbourbeau !