CUDA out of memory in 24G 3090.

GuangxingHan / FCT

Code for CVPR 2022 Oral paper: 'Few-Shot Object Detection with Fully Cross-Transformer'

72 stars 13 forks source link

Open renshouyu opened 1 month ago

renshouyu commented 1 month ago

您好，我想请问一下作者的显卡是多大的啊，我batch_size设为1，运行sh scripts/two_branch_training_pascalvoc_split1_pvt_v2_b2_li.sh的时候还是会爆显存。

GuangxingHan commented 1 month ago

Hi, we used 4 A100 GPUs for training. You can try to change the number of learnable layers (https://github.com/GuangxingHan/FCT/blob/main/configs/fsod/two_branch_training_coco_pvt_v2_b2_li.yaml#L10) or use smaller backbone to save memory.