search

AI4Bharat / IndicXlit

Transliteration models for 21 Indic languages
https://ai4bharat.iitm.ac.in/transliteration
MIT License
67 stars 16 forks source link

Very slow transliteration #20

Open adsk2050 opened 1 year ago

adsk2050 commented 1 year ago

Hello

It is taking a lot of time for doing transliteration on a dataset. What is the time complexity of this library? Is there any way to get this done faster? I don't have a GPU so any other way it can be sped up?

StephennFernandes commented 5 months ago

hey were you able to find a solution for speeding up the process ?

i have been facing similar issues, for sentence that has 47 words the latency is around 600ms, and i need the model to perform inference on text around 2k words at least with 300-600ms latency.

the following is how my code looks like for yielding minimal inference latency. 

translit = XlitEngine(beam_width=1, rescore=False, model_type="transformer", src_script_type = "indic")

@yashmadhani97 @GokulNC hey guys could you please help me out on this ?

currently looking for transliterating docs that contain millions of sentences. (hence inference latency is critical for me)