How to use the aligned vectors? about muse HOT 8 CLOSED

What you could do is start the supervised.py code by providing your source and target embeddings, then just after this line: https://github.com/facebookresearch/MUSE/blob/master/supervised.py#L66
you can modify the mapping attribute of the model to reload your best_mappings.pth. You can check this function to see how to do this:

Something like this should work:

to_reload = torch.from_numpy(torch.load('your_path/best_mappings.pth'))
W = trainer.mapping.weight.data
assert to_reload.size() == W.size()
W.copy_(to_reload.type_as(W))

Then, by simply calling trainer.build_dictionary(), trainer.dico will contain the dictionary you are looking for.

from muse.

Yes, the output vectors are aligned and you can directly use that.

from muse.

@jamsheer2u you need to implement a function that takes a Malayalam word as input, and compute the cosine similarity between this word and all the words in Hindi. Then take the word in Hindi that has the highest cosine with the input Malayalam word.
We will provide a demo ipython notebook to load some word vectors and play with it.

from muse.

from muse.

Hi @jamsheer2u,
This repo is for embeddings only. But we will probably open-source code for unsupervised NMT in the future.

from muse.

How to use best_mappings.pth generated by unsupervised.py on english<---->spanish language pairs to get a dictionary/list of words? @glample

from muse.

thank you!

from muse.

After doing unsupervised training, are the output vectors produced the aligned ones? Or do I need to also use the .pth file to align them.

I'm trying to both build a dictionary and compute the distance between the pairs.

from muse.