netCDF4.stringtochar broken in 1.7.4

[valeriupredoi]

Hi folks, please see the MRE below:

import netCDF4 as nc
import numpy as np

# should work as per https://unidata.github.io/netcdf4-python/#netCDF4.stringtochar
months = np.array(["January", "February", "March", "April"], dtype="S8")
print(nc.stringtochar(months))
# but I get:
# Traceback (most recent call last):
#   File "/home/valeriu/netcdf4_MRE.py", line 5, in <module>
#     print(nc.stringtochar(months))
#           ~~~~~~~~~~~~~~~^^^^^^^^
#   File "src/netCDF4/_netCDF4.pyx", line 6824, in netCDF4._netCDF4.stringtochar
# AttributeError: 'numpy.bytes_' object has no attribute 'encode'. Did you mean: 'decode'?
# with
# Name                     Version          Build                    Channel
# netcdf4                    1.7.4            nompi_py314h4ae7121_101  conda-forge
# everything OK with netCDF4==1.7.3
# on a generic Ubuntu machine

stringtochar is not working anymore in 1.7.4. Apologies for bugging you here, but am not sure how to let the netCDF4-C folks know about it, it's clearly an issue in the C layer, perhaps you can ping them about it? Very many thanks in advance! Cheers 🍻

[ocefpaf]

Those are bytes already. What happens if you do nc.stringtochar(months, encoding="none") instead?

[valeriupredoi]

Those are bytes already. What happens if you do nc.stringtochar(months, encoding="none") instead?

then it works, but then the function docs need to change - since the behaviour changes from 1.7.3 to 1.7.4 - the docs say it ingests a: Input numpy string array with numpy datatype 'SN' or 'UN', where N is the number of characters in each string. I pass it S8 which should be fine without specifying encoding="none" (like it was in 1.7.3)

[jswhit]

The fix to issue #1440 (PR #1443) is the culprit. This fix made stringtochar work correctly for unicode strings - the problem is that unless you specify encoding='ascii' when dtype='S#', it will assume it's getting a unicode string and spit out the error you see.

So, using encoding='ascii' is a short-term workaround which should give you the same behavior as before. That's obviously not satisfying, so I'll look deeper and see if there is a better solution.

[jswhit]

PR #1466 changes the default encoding from utf-8 to ascii if an array of dtype='S' is provided. This should allow the stringtochar/chartostring functions to behave as they did in 1.7.3.

[valeriupredoi]

hi Jeff, very many thanks for looking into this, and for the fix - and thanks to Filipe for the attention given as well - you gents have a good weekend 🍻