Giters

DeepSceneSeg / EfficientPS

PyTorch code for training EfficientPS for Panoptic Segmentation

Home Page:http://panoptic.cs.uni-freiburg.de/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CUDA out of memoey

xuanzhezhao opened this issue · comments

commented

Hi, thanks for your great work! But when I try to run the train/test demo, my GPU memory ran out
(on Tesla V100, 16GB), so I just wonder how should I change the hyperparameters in config file to reduce GPU consumption and still get good performance in the mean while ?

Thanks a lot.

commented

Plus, I can run the test demo successfully while training demo still failed to run due to this CUDA issue.

commented

Update: it works now after I changed 'imgs_per_gpu' and 'workers_per_gpu' to 1 , but also I would greatly appreciate it if you have any other suggestions for this issue. Thanks!

commented

You can reduce either change either the encoder part of the backbone to EfficientNet-b0 or change the random crop resolution to a smaller one or both.
Usually for Cityscapes 12GB is required for 1 image during training