search

gnobitab / RectifiedFlow

Official Implementation of Rectified Flow (ICLR2023 Spotlight)
663 stars 40 forks source link

how to accelerate the training process? #9

Open fikry102 opened 1 year ago

fikry102 commented 1 year ago

python ./main.py --config ./configs/rectified_flow/cifar10_rf_gaussian_ddpmpp.py --eval_folder eval --mode train --workdir ./logs/1_rectified_flow

It seems that the training process need 60w iterations.

It seems that the memory usage of each gpu is not very high during the training process. (4.3G for 24G RTX3090) Is there any way to increase the memory usage and therefore accelerate the training process?

forever208 commented 8 months ago

just increase the batch size and learning rate at the same time

forever208 commented 8 months ago

@fikry102 Hi, how long does it take to train on CIFAR-10? for example, with batch=128, how many steps and how many hours does it cost on a RTX 3090?