Hi!
Thanks for the lib and tutorial, it is very informative.
With respect to finetuning would it be worth quantizing the model first to fp16 or even int8 before beginning training?
As this might lead to better accuracy when compared to quantizing after the model has been finetuned?
Hi! Thanks for the lib and tutorial, it is very informative.
With respect to finetuning would it be worth quantizing the model first to fp16 or even int8 before beginning training? As this might lead to better accuracy when compared to quantizing after the model has been finetuned?
Thanks