Create string ndarray from ndarray of int 'flags' using list of string meanings

Question

I'm trying to go from an ndarray of integer 'flags':

array([[1, 3, 2],
       [2, 0, 3],
       [3, 2, 0],
       [2, 0, 1]])

to an ndarray of strings:

array([['Banana', 'Celery', 'Carrot'],
       ['Carrot', 'Apple', 'Celery'],
       ['Celery', 'Carrot', 'Apple'],
       ['Carrot', 'Apple', 'Banana']],
      dtype='|S6')

Using a list of strings as the mapping of 'flags' to 'meanings':

meanings = ['Apple', 'Banana', 'Carrot', 'Celery']

I've come up with the following:

>>> import numpy as np
>>> meanings = ['Apple', 'Banana', 'Carrot', 'Celery']
>>> flags = np.array([[1,3,2],[2,0,3],[3,2,0],[2,0,1]])
>>> flags
array([[1, 3, 2],
       [2, 0, 3],
       [3, 2, 0],
       [2, 0, 1]])
>>> mapped = np.array([meanings[f] for f in flags.flatten()]).reshape(flags.shape)
>>> mapped
array([['Banana', 'Celery', 'Carrot'],
       ['Carrot', 'Apple', 'Celery'],
       ['Celery', 'Carrot', 'Apple'],
       ['Carrot', 'Apple', 'Banana']],
      dtype='|S6')

This works, but I'm concerned about the efficiency (list comp, flatten, reshape) of the pertinent line when dealing with large ndarrays:

np.array([meanings[f] for f in flags.flatten()]).reshape(flags.shape)

Is there a better/more efficient way of performing a mapping like this?

Answer 1

Fancy indexing is the numpythonic way of doing it:

mapped = meanings[flags]

or the often faster equivalent:

mapped = np.take(meanings, flags)

Answer 2

I think np.vectorize is the way to go, it's also really clear and easy to follow. I haven't tested the following but it should work.

vfunc = np.vectorize(lambda x : meanings[x])

mapped = vfunc(flags)