Convert a numpy matrix to a set of pandas Series

Question

Question: Is there a quick way to convert a 2D Numpy matrix to a set of Pandas Series? For example, a (100 x5) ndarray, to 5 series with 100 rows each.

Background: I need to create a pandas dataframe using randomly generated data of different types (float, string, etc). Currently, for float, I create a numpy matrix, for strings, I create an array of strings. Then I combine all of these along axis=1 to form a dataframe. This does not preserve the datatypes of each individual column.

To preserve the datatype, I plan to use pandas series. Since creating multiple series of floats will likely be slower than creating a numpy matrix of floats, I was wondering if there was a way to convert a numpy matrix to a set of series.

This question is different from mine in that it asks about converting a numpy matrix into a single series. I require multiple series.

Answer 1

You can convert the matrix of each data type directly to a dataframe and then concatenate the resulting dataframes.

float_df = pd.DataFrame(np.random.rand(500).reshape((-1,5)))
#           0         1         2         3         4
#0   0.561765  0.177957  0.279419  0.332973  0.967186
#1   0.761327  0.323747  0.707742  0.555475  0.680662
#..       ...       ...       ...       ...       ...
#98  0.741207  0.061200  0.142316  0.381168  0.591554
#99  0.417697  0.723469  0.730677  0.538261  0.281296
#
#[100 rows x 5 columns]

pd.concat([float_df, int_df, ...], axis=1)

Answer 2

Making a dataframe from a dict of arrays:

In [571]: df = pd.DataFrame({'a':['one','two','three'], 'b':np.arange(3), 'c':np.ones(3)})
In [572]: df
Out[572]: 
       a  b    c
0    one  0  1.0
1    two  1  1.0
2  three  2  1.0

Note the mixed column dtypes:

In [579]: df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 3 columns):
 #   Column  Non-Null Count  Dtype  
---  ------  --------------  -----  
 0   a       3 non-null      object 
 1   b       3 non-null      int64  
 2   c       3 non-null      float64
dtypes: float64(1), int64(1), object(1)
memory usage: 200.0+ bytes

If we ask for a numpy from that, we get a 2d object dtype array:

In [580]: df.values
Out[580]: 
array([['one', 0, 1.0],
       ['two', 1, 1.0],
       ['three', 2, 1.0]], dtype=object)

Recreating a dataframe, looks the same, but the column dtypes are different:

In [581]: pd.DataFrame(df.values, columns=['a','b','c'])
Out[581]: 
       a  b    c
0    one  0  1.0
1    two  1  1.0
2  three  2  1.0
In [582]: _.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 3 columns):
 #   Column  Non-Null Count  Dtype 
---  ------  --------------  ----- 
 0   a       3 non-null      object
 1   b       3 non-null      object
 2   c       3 non-null      object
dtypes: object(3)
memory usage: 200.0+ bytes

But a structured array does preserve column dtpes:

In [587]: df.to_records(index=False)
Out[587]: 
rec.array([('one', 0, 1.), ('two', 1, 1.), ('three', 2, 1.)],
          dtype=[('a', 'O'), ('b', '<i8'), ('c', '<f8')])
In [588]: pd.DataFrame(_)
Out[588]: 
       a  b    c
0    one  0  1.0
1    two  1  1.0
2  three  2  1.0
In [589]: _.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 3 columns):
 #   Column  Non-Null Count  Dtype  
---  ------  --------------  -----  
 0   a       3 non-null      object 
 1   b       3 non-null      int64  
 2   c       3 non-null      float64
dtypes: float64(1), int64(1), object(1)
memory usage: 200.0+ bytes