search

epfml / sent2vec

General purpose unsupervised sentence representations
Other
1.19k stars 256 forks source link

Getting embeddings in real time #6

Closed shengyehchen closed 7 years ago

shengyehchen commented 7 years ago

I followed the instructions in get_sentence_embeddings_from_pre-trained_models.ipynb to get embeddings of sentences, and I found that every time I call the method get_sentence_embeddings(), it reloads the parameters of the model once, making loading sentence embedding several times a time-consuming stuff.

Is there any solutions to make it more efficient?

mpagli commented 7 years ago

The solution is to give huge batches, which is a not perfect but good enough solution when you have all your sentences upfront. For the real time scenario, one could implement a server to serve those embeddings. We did not invest some time in this yet, but this would be a great asset to have.

shengyehchen commented 7 years ago

Indeed, I am trying to use it in the real time scenario, so it would be wonderful if someone implements an interface which separates the loading model & the getting embeddings part!

martinjaggi commented 7 years ago

this concerns a use-case and code part we fully share with fasttext. it would therefore be best to ask the fasttext community directly, or the python API there. please feel free to re-open this issue here once there is any update!

martinjaggi commented 6 years ago

has been solved in #17