Closed 618QRC closed 1 year ago
Hey,
Which part do you mean by training from scratch? But generally all trainings will be mostly bottlenecked by the models forward and backward pass, to language grounding loss or the standard finetuning wont objective wont make a big difference. So hard to say, depending on the GPU type, and local bandwidth, but I would say 2 days is a reasonable estimate on a single larger modern GPU.
Cheers, David
Hey,
Thanks for your reply! So it means training the model for both language grounding pretraining and fine-tuning on a specific task(semantic segmentation) may take 2 days on a single GPU like RXT_3090 or other equal types GPU?
Yes and no :) So both take two days, so it takes 4 days altogether (though the grounding phase is task independent), but you don't have to retrain the preatraining phase after for different downstream tasks.
Hey,
Got it, Thanks for your nice reply.
Have a good day!
Hi we are focusing on the efficiency of 3d models, so we want to know the time performance of this method. BTW, thanks for your talented work, it is fabulous