Pandas: Converting to numeric, creating NaNs when necessary

Say I have a column in a dataframe that has some numbers and some non-numbers

>> df['foo']
0       0.0
1     103.8
2     751.1
3       0.0
4       0.0
5         -
6         -
7       0.0
8         -
9       0.0
Name: foo, Length: 9, dtype: object

How can I convert this column to np.float, and have everything else that is not float convert it to NaN?

When I try:

>> df['foo'].astype(np.float)

or

>> df['foo'].apply(np.float)

I get ValueError: could not convert string to float: -

In pandas 0.17.0 convert_objects raises a warning:

> FutureWarning: convert_objects is deprecated. Use the data-type > specific converters pd.to_datetime, pd.to_timedelta and pd.to_numeric.

You could use pd.to_numeric method and apply it for the dataframe with arg coerce.

df1 = df.apply(pd.to_numeric, args=('coerce',))

or maybe more appropriately:

df1 = df.apply(pd.to_numeric, errors='coerce')

EDIT

The above method is only valid for pandas version >= 0.17.0, from docs what's new in pandas 0.17.0:

> pd.to_numeric is a new function to coerce strings to numbers (possibly with coercion) (GH11133)

Use the convert_objects Series method (and convert_numeric):

In [11]: s
Out[11]: 
0    103.8
1    751.1
2      0.0
3      0.0
4        -
5        -
6      0.0
7        -
8      0.0
dtype: object

In [12]: s.convert_objects(convert_numeric=True)
Out[12]: 
0    103.8
1    751.1
2      0.0
3      0.0
4      NaN
5      NaN
6      0.0
7      NaN
8      0.0
dtype: float64

Note: this is also available as a DataFrame method.

You can simply use pd.to_numeric and setting error to coerce without using apply

df['foo'] = pd.to_numeric(df['foo'], errors='coerce')

First replace all the string values with None, to mark them as missing values and then convert it to float.

df['foo'][df['foo'] == '-'] = None
df['foo'] = df['foo'].astype(float)