CODEWITHSUNDEEP
pythonpandasindexingslice
svenkatesh
svenkatesh

Reputation: 1192

pandas: slice Multindex with many indices

I have a dataframe d with about 100,000,000 rows and 3 columns. It looks something like this: 

import pandas as pd 

In [17]: d = pd.DataFrame({'id': ['a', 'b', 'c', 'd', 'e'], 'val': [1, 2, 3, 4, 5], 'n': [34, 22, 95, 86, 44]}) 

In [18]: d.set_index(['id', 'val'], inplace = True)

I have another dataframe with values of id and val that I want to keep in d. There are around 600,000 combinations of id and val that I want to keep: 

In [20]: keep = pd.DataFrame({'id':['a', 'b'], 'val' : [1, 2]})

I have tried this in the following way: 

In [21]: keep.set_index(['id', 'val'], inplace = True)

In [22]: d.loc[d.index.isin(keep.index), :] 
Out [22]:         
                   n
         id val    
          a  1    34
          b  2    22

This works but seems clunky and is very slow. Is there a better approach here? What is the fastest way to slice on Multindex in pandas?

Upvotes: 4

Views: 151

Answers (3)

cs95
cs95

Reputation: 403050

Use loc to index d using keep:

# d.loc[pd.MultiIndex.from_arrays([keep['id'], keep['val']]))]
d.reindex(pd.MultiIndex.from_arrays([keep['id'], keep['val']])))

         n
id val    
a  1    34
b  2    22

Where,

pd.MultiIndex.from_arrays([keep['id'], keep['val']])

MultiIndex(levels=[['a', 'b'], [1, 2]],
           codes=[[0, 1], [0, 1]],
           names=['id', 'val'])

Generates a MultiIndex from keep that is used to filter.

Another option is to try merging with df.merge:

d.merge(keep, left_index=True, right_on=['id', 'val'])

    n id  val
0  34  a    1
1  22  b    2


d.merge(keep, left_index=True, right_on=['id', 'val']).set_index(['n', 'id'])

       val
n  id     
34 a     1
22 b     2

Or even,

d.join(keep.set_index(['id', 'val']), how='inner')

         n
id val    
a  1    34
b  2    22

Upvotes: 3

BENY
BENY

Reputation: 323366

Using reindex

d.reindex(pd.MultiIndex.from_frame(keep))
Out[151]: 
         n
id val    
a  1    34
b  2    22

Upvotes: 5

piRSquared
piRSquared

Reputation: 294506

loc takes a list of tuples for referencing a MultiIndex

d.loc[[*keep.itertuples(index=False)]]

         n
id val    
a  1    34
b  2    22

A more obnoxious way to do the same thing. (Not actually a recommendation)

d.loc[[*zip(*map(keep.get, keep))]]

         n
id val    
a  1    34
b  2    22

The advantage is 3 less characters. You only have to sacrifice understanding what is going on.

Upvotes: 5

Related Questions