search

zylo117 / Yet-Another-EfficientDet-Pytorch

The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.
GNU Lesser General Public License v3.0
5.21k stars 1.27k forks source link

About GPU memory usage #300

Closed johninn closed 4 years ago

johninn commented 4 years ago

Hello, I got a question about GPU memory usage. I train a model with 46 class and head only false on d7. GPU is Tesla P100 16GB. I got out of memory even for batchsize=1. What GPU do you use for training COCO which has 80 class with efficientdet d7?

zylo117 commented 4 years ago

did you update the latest code? one of previous version could lead to oom.

johninn commented 4 years ago

Yes, I've updated the latest code. I'm trying run head only for 2 epoch, and than add bifpn. With head only false and batchsize=1, it will oom.

zylo117 commented 4 years ago

in that case, try using --optim sgd. if that can't help, you should switch to a small network like d5 or d6.