search

facebookresearch / InferSent

InferSent sentence embeddings
Other
2.28k stars 471 forks source link

Regarding Fasttext file crawl-300d-2M.vec #119

Closed hitzkrieg closed 5 years ago

hitzkrieg commented 5 years ago

On unzipping the Fasttext file crawl-300d-2M.vec.zip , I came across two files: crawl-300d-2M-subword.bin and crawl-300d-2M-subword.vec. However the file crawl-300d-2M.vec expected by demo.ipynb is not there inside this zip leading to error.

nreimers commented 5 years ago

I think the URL in the readme is wrong.

Correct URL: https://dl.fbaipublicfiles.com/fasttext/vectors-english/crawl-300d-2M.vec.zip

HenryDashwood commented 5 years ago

Having now tried using InferSent with both files, I'm also pretty sure the url in the README is wrong. For instance, this is the output of the visualize method when I use crawl-300d-2M.vec image However when I use crawl-300d-2M-subword.vec the central bars are more or less inverted with the obviously less important words receiving greater importance.

saippuakauppias commented 5 years ago

File in readme is: crawl-300d-2M-subword.zip But need crawl-300d-2M.vec.zip

I tested both and -subword give me very bad results.