NumPy dtype issues in genfromtxt(), reads string in as bytestring

I want to read in a standard-ascii csv file into numpy, which consists of floats and strings.

E.g.,

ZINC00043096,C.3,C1,-0.1540,methyl
ZINC00043096,C.3,C2,0.0638,methylene
ZINC00043096,C.3,C4,0.0669,methylene
ZINC00090377,C.3,C7,0.2070,methylene
...

Whatever I tried, the resulting array would look like

E.g.,

all_data = np.genfromtxt(csv_file, dtype=None, delimiter=',')


[(b'ZINC00043096', b'C.3', b'C1', -0.154, b'methyl')
 (b'ZINC00043096', b'C.3', b'C2', 0.0638, b'methylene')
 (b'ZINC00043096', b'C.3', b'C4', 0.0669, b'methylene')

However, I want to save a step for the byte-string conversion and was wondering how I can read in the string columns as regular string directly.

I tried several things from the numpy.genfromtxt() documentation, e.g., dtype='S,S,S,f,S' or dtype='a25,a25,a25,f,a25', but nothing really helped here.

I am afraid, but I think I just don’t understand how the dtype conversion really works…Would be nice if you can give me some hint here!

Thanks

Answer

In Python2.7

array([('ZINC00043096', 'C.3', 'C1', -0.154, 'methyl'),
       ('ZINC00043096', 'C.3', 'C2', 0.0638, 'methylene'),
       ('ZINC00043096', 'C.3', 'C4', 0.0669, 'methylene'),
       ('ZINC00090377', 'C.3', 'C7', 0.207, 'methylene')], 
      dtype=[('f0', 'S12'), ('f1', 'S3'), ('f2', 'S2'), ('f3', '<f8'), ('f4', 'S9')])

in Python3

array([(b'ZINC00043096', b'C.3', b'C1', -0.154, b'methyl'),
       (b'ZINC00043096', b'C.3', b'C2', 0.0638, b'methylene'),
       (b'ZINC00043096', b'C.3', b'C4', 0.0669, b'methylene'),
       (b'ZINC00090377', b'C.3', b'C7', 0.207, b'methylene')], 
      dtype=[('f0', 'S12'), ('f1', 'S3'), ('f2', 'S2'), ('f3', '<f8'), ('f4', 'S9')])

The ‘regular’ strings in Python3 are unicode. But your text file has byte strings. all_data is the same in both cases (136 bytes), but Python3’s way of displaying a byte string is b'C.3', not just ‘C.3’.

What kinds of operations do you plan on doing with these strings? 'ZIN' in all_data['f0'][1] works with the 2.7 version, but in 3 you have to use b'ZIN' in all_data['f0'][1].

Variable/unknown length string/unicode dtype in numpy reminds me that you can specify a unicode string type in the dtype. However this becomes more complicated if you don’t know the lengths of the strings beforehand.

alttype = np.dtype([('f0', 'U12'), ('f1', 'U3'), ('f2', 'U2'), ('f3', '<f8'), ('f4', 'U9')])
all_data_u = np.genfromtxt(csv_file, dtype=alttype, delimiter=',')

producing

array([('ZINC00043096', 'C.3', 'C1', -0.154, 'methyl'),
       ('ZINC00043096', 'C.3', 'C2', 0.0638, 'methylene'),
       ('ZINC00043096', 'C.3', 'C4', 0.0669, 'methylene'),
       ('ZINC00090377', 'C.3', 'C7', 0.207, 'methylene')], 
      dtype=[('f0', '<U12'), ('f1', '<U3'), ('f2', '<U2'), ('f3', '<f8'), ('f4', '<U9')])

In Python2.7 all_data_u displays as

(u'ZINC00043096', u'C.3', u'C1', -0.154, u'methyl')

all_data_u is 448 bytes, because numpy allocates 4 bytes for each unicode character. Each U4 item is 16 bytes long.

---

Changes in v 1.14: https://docs.scipy.org/doc/numpy/release.html#encoding-argument-for-text-io-functions