IndexError: too many indices for array. Numpy array with 42 features not homogeneous

Question

I'm trying to instantiate my test set for classification, loading a dataset with 41 features and 1 label:

import numpy as np

f = open("mydataset")
dataset = np.genfromtxt(f, delimiter=',', dtype=None)

X = dataset[:, 0:40]  # select columns 1 through 41
y = dataset[:, 41]  # select column 42 (the labels)

Since mydataset is not homogeneous (not all elements have the same type), the function genfromtxt creates a 1D array (a list of tuples). So I get this error:

X = dataset[:, 0:40]  # select columns 1 through 41
IndexError: too many indices for array

How can I solve this? Have I to transform the numpy array in 2D (if yes, in which way)? Or have I to use another way to select the right columns?

Thanks

Community · Accepted Answer · 2017-05-23 12:06:44Z

3

You could define a compound dtype:

dt = np.dtype([('values',float,(41,)),('labels','S10')])
data=np.genfromtxt(f, delimiters=',',dtype=dt)
X = data['values']
Y = data['labels']

(not tested because I don't have a sample array this size).

And as I describe in a recent answer, https://stackoverflow.com/a/37126091/901925,

you could convert the dtype=None data to this compound dtype with

data.view(dt)

though that requires that all the numbers be loaded as float (or all as ints). Often CSVs have a mix of float and integer columns, so the numeric fields of a None genfromtxt call will be a mix of types.

Borrowing from that other answer, a general structured array might look like:

In [421]: data=np.array([('label1', 12, 23.2, 232.0), ('label2', 23, 2324.0, 324.0),
       ('label3', 34, 123.0, 2141.0), ('label4', 0, 2.0, 3.0)], 
      dtype=[('f0', '<U10'), ('f1', '<i4'), ('f2', '<f8'), ('f3', '<f8')])

4 fields with different dtypes.

Individual fields can be accessed by name: data['f0'], or a list of names data[['f0','f3']]. But the things you can do with the list of names is limited.

In [426]: data[['f2','f3']]=10
...
ValueError: multi-field assignment is not supported

You can do more if you make a copy, and more if you view it as homogeneous array:

In [427]: d23=data[['f2','f3']].copy()

In [428]: d23
Out[428]: 
array([(23.2, 232.0), (2324.0, 324.0), (123.0, 2141.0), (2.0, 3.0)], 
      dtype=[('f2', '<f8'), ('f3', '<f8')])

In [429]: d23=d23.view((float,(2,)))

In [430]: d23
Out[430]: 
array([[  2.32000000e+01,   2.32000000e+02],
       [  2.32400000e+03,   3.24000000e+02],
       [  1.23000000e+02,   2.14100000e+03],
       [  2.00000000e+00,   3.00000000e+00]])

In [431]: d23+=34

In [432]: d23
Out[432]: 
array([[   57.2,   266. ],
       [ 2358. ,   358. ],
       [  157. ,  2175. ],
       [   36. ,    37. ]])

(changes to d23 do not affect the original data).

edited May 23, 2017 at 12:06

CommunityBot

11 silver badge

answered May 13, 2016 at 20:05

hpaulj

233k14 gold badges260 silver badges392 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

Gil Over a year ago

I have also another problem, the 41 features are also not homogeneous (some of them are string)

hpaulj Over a year ago

I added some examples of accessing a structured array as might be produced with dtype=None.

Gil Over a year ago

Thanks @hpaulj ! Instead what do you think if I used a for-loop to create a list of lists (passing one raw at time) and then convert it into nparray? Works it well with different types of fields?

hpaulj Over a year ago

It may be best if you ask that last one in a new question - with an example.

Collectives™ on Stack Overflow

IndexError: too many indices for array. Numpy array with 42 features not homogeneous

1 Answer 1

4 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related