Surprising behavior of DataFrame.replace

[cpcloud]

This seems a bit surprising:

>>> from string import ascii_lowercase as letters
>>> from pandas import DataFrame
>>> df = DataFrame(list(letters[:4]), columns=['a'])
>>> df
   a
0  a
1  b
2  c
3  d
>>> df.replace({'a': 'b'})
   a
0  a
1  a
2  c
3  d
>>> df.replace({'a': 'c'})
   a
0  a
1  b
2  b
3  d

Does this have to do with padding?

[jreback]

could be a bug

[cpcloud]

also no doc for the first parameter of this method, will submit a pr for that...

[jreback]

also if u have time:

I think we need clarification in the main docs on usage of when to use

filter, replace, select, update, lookup

I also don't think there is an example of filter anywhere

[cpcloud]

np, not sure if will be done 2day, trying to work out the regex replace on frames and just wanted to note this strangeness when i was looking at one of the simplest possible cases of replacement in an object block

[cpcloud]

@jreback A quick glance at indexing.rst shows docs for select and lookup (they could be fleshed out a bit), but nothing for the other 3. Will add the 3 and flesh the other 2 out.

[jreback]

@cpcloud I think replace is somewhere
a complete example

[cpcloud]

ah yes, grin '\.replace' doc shows it

[jankatins]

Wow, this was worrying, but fortunately df.a.replace({'a': 'b'}) and df.replace({"a":{"a":"b"}}) works as intended...

[cpcloud]

@jreback seems the magic is already in _interpolate, I will leave it out in replace though.

[cpcloud]

@jreback The "bug" here is that when you call

df.replace({'a': 'b'})

the column 'a' replaces the value at index 'b' with the previous non-null value because the fill method default is pad. This seems a bit weird. Can fix if u want. I would change to "search the whole frame for occurrences of 'a' and replace with 'b'.

[jreback]

you want this equivalent to

df.replace('a','b')

?

[cpcloud]

yes, is that ok?

[jreback]

let me take a look a little later

was always fuzzy on this anyhow

are there tests for this usage anyhow ?
do they break?

[cpcloud]

There are tests. However, the replace method doesn't actually do anything so they don't break. Here's a pared down example from the test_replace_interpolate method of TestDataFrame from pandas/tests/test_frame.py:

# get the test frame
from pandas.tests.test_frame import _tsframe as df
from numpy.testing import assert_array_equal
res = df.replace({'A': nan}, method='pad', axis=1)
assert_array_equal(df, res)  # replace == identity function here

[jreback]

I think it makes sense to not have automatic interpolation by default

so, if say df.replace({ 'a' : 'b' }) would be translated to df.replace('a','b'), IOW, only allow None for the value if to_replace is a dict/Series (right now this will interpolate)

then make _interpolate an explicit method interpolate. Even though this is an API change, I doubt this is expected behavor

@wesm, @y-p any thoughts?

@cpcloud why don't you put up the PR when you are ready (as this would be an easy change anyhow)

[cpcloud]

okay. kicking tires a bit right now (added 12 new tests so far); adding the regex functionality is proving to be quite a bit more involved than i thought, but i have a much better understanding (although not complete) of DataFrame internals.

[cpcloud]

@jreback @y-p Can I address this issue in the regex pr? I think it will create fewer merge conflicts if I just do this all at once.

[jreback]

yes, I would

[cpcloud]

just realized the axis parameter in replace never gets used if the call to _interpolate is nixed...remove the parameter?

[jreback]

@cploud close this too?

[cpcloud]

Not just yet. Need to submit 0.12 pr.