datatype parameter in load_word2vec_format doesn't work as expected

[jayantj]

Description

Using datatype=np.float64 in a KeyedVectors.load_word2vec_call doesn't work as expected, the loaded floats seem to lose precision. The datatype for syn0 is still float64 though, so it seems that they are cast to float32 first while loading, then cast to float64 when creating the array.

Steps/Code/Corpus to Reproduce

Using this file -
test.kv.txt

from gensim.models.keyedvectors import KeyedVectors
import numpy as np

kv = KeyedVectors.load_word2vec_format('test.kv.txt', datatype=np.float64)
print(kv['horse.n.01'][0] == -0.0008546282343595379)
# False
print(kv['horse.n.01'].dtype)
# float64

Expected Results

print(kv['horse.n.01'][0] == -0.0008546282343595379)
# True

Actual Results

print(kv['horse.n.01'][0] == -0.0008546282343595379)
# False

Looking at the code and making a quick hack here, changing..

word, weights = parts[0], [REAL(x) for x in parts[1:]]

to..

word, weights = parts[0], [datatype(x) for x in parts[1:]]

..leads to the correct result. However, I imagine there are other cases to be covered as well.