pd vs np for detecting null

jupyter-console.txt

In [16]: data_pd = pd.Series(np.random.rand(10000000))

In [17]: data_pd[data_pd < .2] = np.nan

In [18]: %timeit np.isfinite(data_pd)
10 loops, best of 3: 27.3 ms per loop

In [19]: %timeit pd.notnull(data_pd)
100 loops, best of 3: 11.2 ms per loop

In [20]: # check for equality

In [21]: (pd.notnull(data_pd).values == np.isfinite(data_pd)).all()
Out[21]: True

In [22]: data_np = np.random.rand(10000000)

In [23]: data_np[data_np < .2] = np.nan

In [24]: %timeit pd.notnull(data_np)
100 loops, best of 3: 11.2 ms per loop

In [25]: %timeit np.isfinite(data_np)
10 loops, best of 3: 24.6 ms per loop

In [26]: (pd.notnull(data_np) == np.isfinite(data_np)).all()
Out[26]: True

Wow, nice experiment, good result to know!

You're doing the operations on the pandas object. The reverse is actually true if you do the operations on the ndarray, and it's not even close (numpy is ~ 25x faster)

obj = pd.Series([4, np.nan, 7, np.nan, -3, 2])
%timeit np.isnan(obj).values # Note this one will perform similarly to the Pandas case
%timeit obj.isnull()
%timeit np.isnan(obj.values)

30.5 µs ± 2.87 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
32.9 µs ± 2.53 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
1.15 µs ± 36.3 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

obj = pd.Series([4, 6.5, 7, 3.25, -3, 2])
%timeit obj.div(obj.iloc[::-1])
%timeit obj.values / obj.iloc[::-1].values # Uses the numpy div ufunc

342 µs ± 15.5 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
33.6 µs ± 494 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

%timeit obj.sum()
%timeit np.sum(obj.values)

49 µs ± 867 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
2.7 µs ± 66.5 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

%timeit obj.mean()
%timeit np.mean(obj.values)

22.8 µs ± 885 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
5.46 µs ± 52.1 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

df = pd.DataFrame([[1.4, np.nan], [7.1, -4.5], [np.nan, np.nan], [0.75, -1.3]], index=['a','b','c','d'], columns=['one','two'])
%timeit np.nansum(df.values, axis=0) # Note I switched up the order to make it clear it's not a display fluke
%timeit df.sum()

13.3 µs ± 180 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
94.8 µs ± 1.78 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

And here's your own test case:

data_np = np.random.rand(10000000)
data_np[data_np < .2] = np.nan
data_pd = pd.Series(data_np)
# On the Pandas object
%timeit np.isfinite(data_pd.values)
%timeit ~np.isnan(data_pd.values)
%timeit pd.notnull(data_pd)
# On the Numpy object
%timeit np.isfinite(data_np)
%timeit ~np.isnan(data_np)
%timeit pd.notnull(data_np)

10.6 ms ± 517 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
14.5 ms ± 256 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
14.5 ms ± 669 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
10.4 ms ± 583 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
14.5 ms ± 179 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
14.8 ms ± 307 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Running:

• Python 3.6.4 |Anaconda custom (64-bit)| (default, Jan 16 2018, 10:22:32) [MSC v.1900 64 bit (AMD64)] on win32
• numpy=1.13.3=py36h4a99626_2
• pandas=0.22.0=py36h6538335_0
• anaconda-client=1.6.9=py36_0
• anaconda=custom=py36h363777c_0
• anaconda-navigator=1.7.0=py36_0
• anaconda-project=0.8.2=py36hfad2e28_0