BUG: Interpolate should be more careful with dtypes

[TomAugspurger]

Originally reported as a comment in #6281 Reposting the example from there:

If you start out with a DataFrame like

In [14]: df = DataFrame({'A': [1, 2, np.nan, 4, 5, np.nan, 7],
   ....:                 'C': [1, 2, 3, 5, 8, 13, 21]})

In [15]: df.dtypes
Out[15]: 
A    float64
C      int64
dtype: object

So one float, one int, with no NaNs in the int. We apply the interpolation method to each block, so it gets applied to the int block/column. The interpolation will generally return a float dtype.

In [18]: df.interpolate(method='cubic', downcast=None).dtypes
Out[18]: 
A    float64
C    float64
dtype: object

[jreback]

@TomAugspurger oddly it seems the ref branch #6291 is not accessable.....maybe push your local branch and open a new PR?

[jreback]

@TomAugspurger I emailed github....very odd that that PR can't be accessed

[TomAugspurger]

Very strange. I'm looking through my reflog right now and I'm a bit worried that I pushed from the wrong branch.

ccff293 HEAD@{0}: checkout: moving from master to master
ccff293 HEAD@{1}: rebase -i (finish): returning to refs/heads/master
ccff293 HEAD@{2}: rebase -i (pick): add missing comma from GH 6330
7fd87f3 HEAD@{3}: rebase -i (fixup): check in interp_with_fill too
d3e4f0e HEAD@{4}: rebase -i (start): checkout upstream/master
20058e6 HEAD@{5}: rebase -i (finish): returning to refs/heads/master
20058e6 HEAD@{6}: rebase -i (pick): add missing comma from GH 6330
32d100a HEAD@{7}: rebase -i (pick): use is_ instead of isinstance
d3e4f0e HEAD@{8}: rebase -i (pick): check in interp_with_fill too
3e14c73 HEAD@{9}: rebase -i (pick): BUG: preserve dtypes in interpolate
186491f HEAD@{10}: rebase -i (start): checkout upstream/master
690f97e HEAD@{11}: commit: use is_ instead of isinstance
3023077 HEAD@{12}: reset: moving to origin/interpolate-ignore-good

The @{12} line is what worries me. I did the work across two computers. My latest update was from the other computer so I was reseting to what I had pushed. But I think I reset my local master branch to my remote PR branch. I didn't realize that until made those last couple changes, pushed everything, and tried to change to master (which I was already on). I have no idea if that's related to the github issue though. I'll make a new PR for now.

[TomAugspurger]

Lovely. I'm unable to make a new PR against pandas. I think I messed something up badly. I'll see what I can do.

@jreback are you able to see my branch at https://github.com/TomAugspurger/pandas/tree/interpolate-ignore-good3? I guess if you want to get this merged before it falls too far behind master you could clone it and do a PR.

[jreback]

yes....

pull it down as a new branch - something like

git checkout -b new_branch
git fetch yourremotename/interpolate-ignore-good3

I think you can then just push it up as a new branch