Error with binary fasttext embeddings about muse HOT 4 CLOSED

For now this is not possible, but this is definitely something we were planning to do at some point. Applying the rotation matrix to every word embedding and ngram embedding in the fastText binary (i.e. the entire input matrix) should do the trick. I don't know if the fastText API allows it. It looks like the API has a "save_model" functionality, but I don't know if you can modify the components of a model easily.

Unfortunately I won't have time to look at this anytime soon but if you want to try I can help you integrating it into MUSE.

from muse.

Hi,

There was indeed a small bug (which appeared when fastText is trained with ngrams and not without ngrams like in a word2vec setting). This commit was designed to handle the bug:
6e0b460

Are you using the latest version? Otherwise, maybe check that the dimension of your embeddings is 300. Otherwise, can you print and copy here the value of (len(words), params.emb_dim) and of embeddings.size()?

from muse.

Yes, thank you, it trains now. I want to get the OOV words translated (there is issue #13 , but I cannot get embeddings described there).

Is there a way to get out binaries as well ? The output is torch.save(), but I was wondering if it is possible to get the binaries that fasttext(python library or gensim version) ? Is there way to transform it into bin format ?

from muse.

I do not have time as well right now, but maybe if I have time over the summer.

from muse.