Error with pre-trained word embeddings

[nusnlp]

Error with pre-trained word embeddings

[michellegiang]

Hi,

I also post another issue at below link when I train with my data.

https://github.com/facebookresearch/fairseq-py/issues/129

[shamilcm]

Hi, the current issue seems to be with our fork of fairseq-py, so you can close the issue you opened here: facebookresearch/fairseq-py#129 and re-post the issue here.

[shamilcm]

The issue was due to some hardcoded paths in arguments.
It is now fixed here: shamilcm/fairseq-py@ceb2f12
Can you retry with this?

Regarding the training issue, can you close it at facebookresearch/fairseq-py#129 and open a new issue here. I will take a look at it.

Also what data is it trained on? Is it a very small training data with fewer than 30000 words in the vocabulary?

[michellegiang]

hi, the training data is Lang-8 Learner Corpus of English v1.0 and NUCLE

[michellegiang]

Hi Shamil,

So I need to delete the software/fairseq-py, download the new one and reinstall it right ?

Regards,
Viet Anh

[shamilcm]

If you installed fairseq-py using setup.py, pull the new changes and run it again. Otherwise, you just need to just pull the changes. The change is done in only one file: fairseq/utils.py

[michellegiang]

Thank Shamil. If I already installed fairseq-py, could I just copy your new utils to replace the old utils ?

The reason is that I used your version of fairseq-py with the new version of PyTorch and I had some trouble with setup.py build and I need to apply some manual patch. (The version of PyTorch in your original read me has some problems, thus I need to use the newest version of PyTorch)

https://github.com/facebookresearch/fairseq-py/issues/120

[shamilcm]

If you installed by python setup.py develop, just getting utils.py should work.

[shamilcm]

Closing this issue as it has been resolved.