Increasing the speed of translation

(Milton Li) #1

I am currently hosting a trained model online and trying to translate stuff with the model on-the-go. I have lowered the beam_size, but I realized a lot of time is wasted on initializing the model over and over again. Does anyone know:

  1. How else can I increase the speed of translation?
  2. How can I configure the model or translate.lua to stay active and “listen” for calls for translation so that the model need not spend time being terminated and reinitialized subsequently?


(Guillaume Klein) #2

Are you translating on CPU or GPU?

You probably you want to use one of the provided translation servers.

(Milton Li) #3

I’m using GPU to translate. I’ve tried using the translation servers and it works perfectly, thank you.

(miguel canals) #4


What kind of client/server setup are you using? Is a single sentence task or a multiple sentence task?

REST API server is pretty fast for a sentence by sentence task, but not sure what happens with a huge model.

Have a nice day
miguel canals