ENH: Added DataFrame.round and associated tests

[roblevy]

I've found myself doing a lot of DataFrame.to_latex because I'm using pandas to write an academic paper.

I'm constantly messing about with the number of decimal places displayed by doing np.round(df, 2) so thought this flexible round, with different numbers of decimals per column, should be part of the DataFrame API (I'm surprised there isn't already such a piece of functionality.)

Here is an example:

In [9]: df = pd.DataFrame(np.random.random([10, 3]), columns=['a', 'b', 'c']) In [10]: df Out[10]: a b c 0 0.761651 0.430963 0.440312 1 0.094071 0.242381 0.149731 2 0.620050 0.462600 0.194143 3 0.614627 0.692106 0.176523 4 0.215396 0.888180 0.380283 5 0.492990 0.200268 0.067020 6 0.804531 0.816366 0.065751 7 0.751224 0.037474 0.884083 8 0.994758 0.450143 0.808945 9 0.373180 0.537589 0.809112 In [11]: df.round(dict(b=2, c=4)) Out[11]: a b c 0 0.761651 0.43 0.4403 1 0.094071 0.24 0.1497 2 0.620050 0.46 0.1941 3 0.614627 0.69 0.1765 4 0.215396 0.89 0.3803 5 0.492990 0.20 0.0670 6 0.804531 0.82 0.0658 7 0.751224 0.04 0.8841 8 0.994758 0.45 0.8089 9 0.373180 0.54 0.8091 

You can also round by column number:

In [12]: df.round([1, 2, 3]) Out[12]: a b c 0 0.8 0.43 0.440 1 0.1 0.24 0.150 2 0.6 0.46 0.194 3 0.6 0.69 0.177 4 0.2 0.89 0.380 5 0.5 0.20 0.067 6 0.8 0.82 0.066 7 0.8 0.04 0.884 8 1.0 0.45 0.809 9 0.4 0.54 0.809 

and any columns which are not explicitly rounded are unaffected:

In [13]: df.round([1]) Out[13]: a b c 0 0.8 0.430963 0.440312 1 0.1 0.242381 0.149731 2 0.6 0.462600 0.194143 3 0.6 0.692106 0.176523 4 0.2 0.888180 0.380283 5 0.5 0.200268 0.067020 6 0.8 0.816366 0.065751 7 0.8 0.037474 0.884083 8 1.0 0.450143 0.808945 9 0.4 0.537589 0.809112 

Non-integer values raise a TypeError, as might be expected:

In [15]: df.round({'a':1.2}) --------------------------------------------------------------------------- TypeError Traceback (most recent call last) <ipython-input-15-6f51d3fd917d> in <module>() ----> 1 df.round({'a':1.2}) /home/rob/Dropbox/PhD/pandas/pandas/core/frame.py in round(self, places) 1467 1468 if isinstance(places, dict): -> 1469 new_cols = [col for col in _dict_round(self, places)] 1470 else: 1471 new_cols = [col for col in _list_round(self, places)] /home/rob/Dropbox/PhD/pandas/pandas/core/frame.py in _dict_round(df, places) 1455 for col in df: 1456 try: -> 1457 yield np.round(df[col], places[col]) 1458 except KeyError: 1459 yield df[col] /usr/local/lib/python2.7/dist-packages/numpy/core/fromnumeric.pyc in round_(a, decimals, out) 2646 except AttributeError: 2647 return _wrapit(a, 'round', decimals, out) -> 2648 return round(decimals, out) 2649 2650 /home/rob/Dropbox/PhD/pandas/pandas/core/series.pyc in round(self, decimals, out) 1209 1210 """ -> 1211 result = _values_from_object(self).round(decimals, out=out) 1212 if out is None: 1213 result = self._constructor(result, TypeError: integer argument expected, got float 

[jreback]

is there a reason you are not simply using the options display.float_format or display.precision?

[jreback]

In [1]: np.random.seed(1234) In [2]: df = pd.DataFrame(np.random.random([10, 3]), columns=['a', 'b', 'c']) In [4]: df Out[4]: a b c 0 0.191519 0.622109 0.437728 1 0.785359 0.779976 0.272593 2 0.276464 0.801872 0.958139 3 0.875933 0.357817 0.500995 4 0.683463 0.712702 0.370251 5 0.561196 0.503083 0.013768 6 0.772827 0.882641 0.364886 7 0.615396 0.075381 0.368824 8 0.933140 0.651378 0.397203 9 0.788730 0.316836 0.568099 In [7]: pd.set_option('float_format',lambda x: "%.3f" % x) In [8]: df Out[8]: a b c 0 0.192 0.622 0.438 1 0.785 0.780 0.273 2 0.276 0.802 0.958 3 0.876 0.358 0.501 4 0.683 0.713 0.370 5 0.561 0.503 0.014 6 0.773 0.883 0.365 7 0.615 0.075 0.369 8 0.933 0.651 0.397 9 0.789 0.317 0.568 In [9]: pd.set_option('precision',3) In [10]: df Out[10]: a b c 0 0.192 0.622 0.438 1 0.785 0.780 0.273 2 0.276 0.802 0.958 3 0.876 0.358 0.501 4 0.683 0.713 0.370 5 0.561 0.503 0.014 6 0.773 0.883 0.365 7 0.615 0.075 0.369 8 0.933 0.651 0.397 9 0.789 0.317 0.568 

[shoyer]

I think this is a great idea and would be a useful API addition.

A few comments on the design:

1. We should support supplying a single scalar value, e.g. df.round(2).
2. This should also be a Series method (for scalar values). Thus the logic should perhaps live in pandas/core/generic.py.
3. df.round([1]) should be a ValueError -- if you are referencing columns by position, you should need to supply a list with the same length as the full dataframe. It's not obvious what the single element list refers to (e.g., the first or last column?), so we should resist the temptation to guess.

@jreback float_format and precision don't let you customize the precision by column. Also, this could be useful for other things that these options don't effect, e.g., rounding prior to exporting the data with to_csv.

[jreback]

I disagree entirely this is very duplicative
why don't we just override each method in numpy?

this is exactly what apply(np.round, axis=1) is for

[shoyer]

Numpy arrays do have a round method, and DataFrame currently wraps almost every NumPy methods. So I think it would be entirely appropriate to add round. IMO it's more surprising that it's missing.

[jreback]

np.round ???

[jreback]

ufuncs are there for a reason

[jorisvandenbossche]

I also think this would be a nice addition.

First, I think it should be clear that rounding and precision display can be different things. There are enough cases where you don't want to just change how the output looks, but where you want 'real' rounding.

There is indeed np.round(df). But when you have heterogeneous dtypes in your dataframe (eg one column with strings), this already does not work.
Plus, the convenience of being able to specify different precisions per column would possibly be nice.

[roblevy]

@jreback display.float_format and display.precision all round each column to the same precision. This functionality allows each column to be rounded to a different precision. This is also why apply(np.round, axis=1) is not the same piece of functionality.

[jreback]

@roblevy I get why you want this functionaility and I suppose a small expansion of the API is ok, but it has become a creeping API :)

My main concern is that methods should not have any real notion of selection (parameterisation is ok, e.g. .drop_duplicates though).

as @shoyer points out this should take a Series/dict only & certainly should be in pandas/core/generic.py

e.g.

df.round(Series([1,2],list('ab')) or df.round(dict(a=1,b=2))

rather than a positional indicator

[roblevy]

Glad to have this accepted in principle, @jreback . Can I make one last effort to convince you that a list input is a good idea?

If, as @shoyer points out, the number of elements in the list match the number of columns in the DataFrame then there's no ambiguity and I would "expect" this to just work. Then columns which are not to be rounded can have a None element associated with them.

Yay? Or still nay?

[shoyer]

I also don't see a strong need for list like input. A dictionary is more flexible and probably more readable in practice, too

[jreback]

yeh list-like just cause confusion. Only should accept dict-like.

[jreback]

@roblevy can you update / fix so passing

[roblevy]

Didn't realise this before, but np.round dispatches to the .round method of whatever you pass it, so I've updated the signature to match what np.round expects.

Thus, the signature is now:

df.round(decimals, out=None) 

where out raises a NotImplementedError. We also have to allow decimals to be an integer (or np.integer) because doing np.round(df, 2) simply dispatches to df.round(2).

[roblevy]

Ok. @jreback looks like we're good to go.

[jorisvandenbossche]

some leftover of rebase?

[jreback]

I would move this to enhancements.

[jorisvandenbossche]

@roblevy I don't think the out kwarg is needed here. We also don't provide this in eg DataFrame.mean to conform to numpy

[shoyer]

@jorisvandenbossche actually, we allow (and ignore) **kwargs on DataFrame.mean, specifically for this reason (numpy compat). Otherwise np.mean does not work on a DataFrame.

So I think this is probably the right approach here.

[shoyer]

decimals should default to 0, like NumPy.

[jreback]

just add **kwargs for compat with numpy

[shoyer]

Sadly numpy needs a positional argument for out with np.round. So I think we should stick with this signature, just removing it from the docstring

[jorisvandenbossche]

@shoyer ah yes, OK, but then it is just not explicitely needed (not in the docstring/not raise an NotImplementedError)

[jreback]

remove the out parameter (we are just ignoring it)

[jreback]

put a blank line between tests

[jreback]

@roblevy don't be scared off :) as you just got a lot of comments.

[jorisvandenbossche]

I don't think this is the right place (it is not a binary operator). But not sure what the good place is ..

Maybe 'Computations / Descriptive Stats'

[jorisvandenbossche]

To be consistent, I would the take the same approach as in other numpy-like methods, which is ignoring out instead of raising an error.

[jreback]

this should be an int dtype series. I think you have to require >= 0. I suppose you could ignore nans as well. I am not sure what np.round would do with these cases, so pls add some tests for validation. If the errors are obtuse, then may need to catch and report a better message.

[jreback]

pls also add a section to the docs http://pandas-docs.github.io/pandas-docs-travis/options.html#number-formatting (or if you have a better idea where then pls report).

[jreback]

can you rebase / update according to comments

[roblevy]

I'm going to push back on handling nan here, because any Series with a nan value becomes a float anyway so numpy returns a sensible error that an integer argument is expected. That ok @jreback?

[shoyer]

You can let numpy handle the error processing, but please add a unit test to verify that an appropriate error is raised.

On Tue, Sep 1, 2015 at 9:21 AM, Rob Levy notifications@github.com wrote:

I'm going to push back on handling nan here, because any Series with a nan value becomes a float anyway so numpy returns a sensible error that an integer argument is expected. That ok @jreback?

Reply to this email directly or view it on GitHub:
#10568 (comment)

[jreback]

you can use :meth:DataFrame.round`` for these

[jreback]

@roblevy ok, pls add the tests @shoyer suggests, minor doc comments. pls squash. ping when green.

[roblevy]

@jreback Good to go. I've added the tests as requested, and updated the docs as requested. Tests are passing!

[jreback]

@roblevy thanks. nice change!

[roblevy]

YAAAY!!! Excellent. I feel very proud.