word2vec model .bin

,
1

I've downloaded a pre-trained word2vec Twitter model and i was wondering if it's possible to load that instead of a sense2vec vector or en_core_web_(any_size)?
Tried passing it as input and got the following error:

  File "/home/USER/anaconda3/envs/py-earthquakes-new/lib/python3.8/site-packages/spacy/util.py", line 253, in get_model_meta
raise IOError(Errors.E053.format(path=meta_path))
OSError: [E053] Could not read meta.json from classification/word2vec_twitter_model.bin/meta.json

Thanks in advance!

2

Which version of spaCy are you using? I'm assuming some version of spaCy 2?

To use an external vectors format, you'll first have to convert it to a spaCy model, cf. https://spacy.io/usage/vectors-similarity#converting. Your model will have to be compatible with one of the formats supported by init-model though...

1 Like
3 
from pathlib import Path

from gensim.models import KeyedVectors
from gensim.test.utils import get_tmpfile
import typer

import spacy
from spacy.cli.init_model import init_model


def unpack_w2v_binary(w2v_path: str):
    txt_vec_file = get_tmpfile("vec.txt")
    KeyedVectors.load_word2vec_format(w2v_path, binary=True).save_word2vec_format(
        txt_vec_file, binary=False
    )
    return txt_vec_file


def build_w2v_spacy_model(
    lang: str, output_dir: str, w2v_path: str, model_name: str
) -> spacy.language.Language:
    """Thin wrapper around SpaCy's init_model_cli command"""
    init_model(
        lang,
        Path(output_dir),
        vectors_loc=unpack_w2v_binary(w2v_path),
        model_name=model_name,
    )


if __name__ == "__main__":
    typer.run(init_model)

^ That works to basically load a spacy model with custom vectors in a bin file. I think there are some caveats with vocab compatibility, but that's a start.