BUG: DataFrame.clip sets all values to the lower_bound

[dhirschfeld]

The following unit-test fails:

def test_dataframe_clip(lb=-1, ub=1):
    df = pd.DataFrame(np.random.randn(1000,2))
    lb_mask = df.values <= lb
    ub_mask = df.values >= ub
    mask = ~lb_mask & ~ub_mask
    clipped_df = df.clip(lb, ub)
    assert (clipped_df.values[lb_mask] == lb).all()
    assert (clipped_df.values[ub_mask] == ub).all()
    assert (clipped_df.valus[mask] == df.values[mask]).all()
#

[jreback]

    def clip(self, upper=None, lower=None):
        """
        Trim values at input threshold(s)

        Parameters
        ----------
        lower : float, default None
        upper : float, default None

        Returns
        -------
        clipped : DataFrame

try this df.clip(upper=ub,lower=lb)
the arguments are 'backwards'

[dhirschfeld]

Wow, that's confusing!

Even if there was a good argument for going against the numpy convention, it's not even consistent within pandas as Series inherits it's method from numpy.

Any chance this can be changed? I know it would break code already out there, but I'm betting there's incorrect code out there already where someone has assumed the numpy convention or tried to write generic Series/DataFrame code using clip.

np.clip(a, a_min, a_max, out=None)
pd.Series.clip(self, lower=None, upper=None, out=None)
pd.DataFrame.clip(self, upper=None, lower=None)

[jreback]

agreed...should prob be changed...

you can also accomplish what you are after like this (but prob more confusing)...though same syntax works on series

df.where(df>-1,-1).where(df<1,1)

[jreback]

closed in #2750

[l736x]

Why not the same treatment for Series? (It just happened that I had a bug due to this)

[jreback]

what version?

[l736x]

Sorry, I missed that issue. I'm on 0.12, I just checked master and there is no sign of:

lower, upper = min(lower, upper), max(lower, upper)

in series.clip
If it's already fixed by 4324, forget it. Sorry for the confusion.