FastText wrapper returns inconsistent dtypes

[mcobzarenco]

Description

gensim.models.wrappers.FastText returns inconsistent dtypes.

Steps/Code/Corpus to Reproduce

from gensim.models.wrappers import FastText
embeds = FastText.load_fasttext_format(...)

For an existing word:

embeds['the'].dtype == dtype('float32')

For an "imputed" word (missing from the vocabulary). The word embedding is computed as the sum of embedding for n-grams:

embeds['ttttt'].dtype == dtype('float64')

The problem in models/wrappers/fasttext.py::FastTextKeyedVectors.word_vec. In the case of a missing word, the zero vector is initialised to be a 64-bit float array to which a bunch of 32-bit embeddings are added to.

Versions

Linux-4.4.0-97-generic-x86_64-with-Ubuntu-16.04-xenial
Python 3.5.2 (default, Nov 17 2016, 17:05:23)
[GCC 5.4.0 20160609]
NumPy 1.13.3
SciPy 0.19.1
gensim 3.0.1
FAST_VERSION 1

[piskvorky]

Nice catch @mcobzarenco ! Thanks.