Slicing a string from a Numpy structured array throws an error

[marnixhoh]

This works:

from numba import njit
import numpy as np

values_dtype = np.dtype([
    ('one', 'U25'),
    ('two', 'f8')
])

def my_test():
    values = np.zeros(2, dtype=values_dtype)
    values['one'][0] = '2020-03-03 00:01:00'
    stuff = values['one'][0][0:10]
    return stuff
result = my_test()
print(result)

But when jitting it, it throws an error:

from numba import njit
import numpy as np

values_dtype = np.dtype([
    ('one', 'U25'),
    ('two', 'f8')
])

@njit
def my_test():
    values = np.zeros(2, dtype=values_dtype)
    values['one'][0] = '2020-03-03 00:01:00'
    stuff = values['one'][0][0:10]
    return stuff
result = my_test()
print(result)

Invalid use of Function() with argument(s) of type(s): ([unichr x 25], slice)

numba: 0.49.1
numpy: 1.18.4
python: 3.7.1

• I am using the latest released version of Numba (most recent is visible in
  the change log (https://github.com/numba/numba/blob/master/CHANGE_LOG).
• I have included below a minimal working reproducer (if you are unsure how
  to write one see http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports).

[marnixhoh]

I just found a work around. If you parse the string from the structured array using str(), it works. However, I am not sure what the performance penalty is for doing so??

from numba import njit
import numpy as np

values_dtype = np.dtype([
    ('one', 'U25'),
    ('two', 'f8')
])

@njit
def my_test():
    values = np.zeros(2, dtype=values_dtype)
    values['one'][0] = '2020-03-03 00:01:00'
    stuff = str(values['one'][0])[0:10]
    return stuff
result = my_test()
print(result)

[marnixhoh]

I also wanted to point out that some string methods do work as expected. E.g. .split():

from numba import njit
import numpy as np

values_dtype = np.dtype([
    ('one', 'U25'),
    ('two', 'f8')
])

@njit
def my_test():
    values = np.zeros(2, dtype=values_dtype)
    values['one'][0] = '2020-03-03 00:01:00'
    stuff = values['one'][0].split(':')
    return stuff
result = my_test()
print(result)

Hope this information is helpful

[stuartarchibald]

Thanks for the report. In the first post #5802 (comment) I think that this won't work as the type of the returned thing will depend on the values in the slice, e.g. [1:3] would produce a [unichr x 2], whereas [1:4] would produce a [unichr x 3].

With respect to your work around #5802 (comment) the performance penalty is likely that a string buffer has to be allocated and the characters copied in, without looking at the machine code I'd guess that it cannot be optimised away as it's a change of internal representation.

As to some method's working, it's because the character array will "look like" a string for the method call if the types can be statically determined.

[stuartarchibald]

Closing this question as it seems to be resolved. Numba now has a discourse forum https://numba.discourse.group/ which is great for questions like this, please do consider posting there in future :) Thanks!