hustvl/QueryInst

Training Memory Usage

Closed this issue · 2 comments

shikunyu8 commented

Thanks for sharing the implementation!

I got memory error with this config https://github.com/hustvl/QueryInst/blob/main/configs/queryinst/queryinst_swin_large_patch4_window7_fpn_300_proposals_crop_mstrain_400-1200_50e_coco.py. I used 8 GPUs each with 32 GB memeory.

What GPUs did you use to conduct your experiments?

image

vealocia commented

For swin-transformer backbone, we follow the official implementation and use torch.checkpoint for memory reduction. Please use the following commands:

./tools/dist_train.sh configs/queryinst/queryinst_swin_large_patch4_window7_fpn_300_proposals_crop_mstrain_400-1200_50e_coco.py 8 --cfg-options model.pretrained=<PRETRAIN_MODEL> model.backbone.use_checkpoint=True

pretrain model can be found here
Hope this helps!

  • 👍1
shikunyu8 commented

Oh, I see. Thanks for the reply!
I'm closing this issue.