Forced conversion of non-numeric numpy arrays with

问题:

Consider the array

x = np.array(['1', '2', 'a'])

Tying to convert to a float array raises an exception

x.astype(np.float)
ValueError: could not convert string to float: a

Does numpy provide any efficient way to coerce this into a numeric array, replacing non-numeric values with something like NAN?

Alternatively, is there an efficient numpy function equivalent to np.isnan, but which also tests for non-numeric elements like letters?

回答1:

You can convert an array of strings into an array of floats (with NaNs) using np.genfromtxt:

In [83]: np.set_printoptions(precision=3, suppress=True)

In [84]: np.genfromtxt(np.array(['1','2','3.14','1e-3','b','nan','inf','-inf']))
Out[84]: array([ 1.   ,  2.   ,  3.14 ,  0.001,    nan,    nan,    inf,   -inf])

---

Here is a way to identify "numeric" strings:

In [34]: x
Out[34]: 
array(['1', '2', 'a'], 
      dtype='|S1')

In [35]: x.astype('unicode')
Out[35]: 
array([u'1', u'2', u'a'], 
      dtype='<U1')

In [36]: np.char.isnumeric(x.astype('unicode'))
Out[36]: array([ True,  True, False], dtype=bool)

Note that "numeric" means a unicode that contains only digit characters -- that is, characters that have the Unicode numeric value property. It does not include the decimal point. So u'1.3' is not considered "numeric".

回答2:

If you happen to be using pandas as well you could use the pd.to_numeric() method:

In [1]: import numpy as np

In [2]: import pandas as pd

In [3]: x = np.array(['1', '2', 'a'])

In [4]: pd.to_numeric(x, errors='coerce')
Out[4]: array([  1.,   2.,  nan])