search

PeizeSun / TransTrack

Multiple Object Tracking with Transformer
MIT License
629 stars 109 forks source link

CUDA out of memory #51

Closed BrainPotter closed 2 years ago

BrainPotter commented 2 years ago

Hi, thank you for your great work. My OS is ubuntu16.04, which has 4 NVIDIA Geforce RTX 2080, When I run the program sh track_exps/crowdhuman_mot_trainhalf.sh, it gave the error CUDA out of memory. the crowdhuman_mot_trainhalf.sh is set as: python3 -m torch.distributed.launch --nproc_per_node=4 --use_env main_track.py --output_dir ./output/crowdhuman_mot_trainhalf --dataset_file mot --coco_path mot --batch_size 2 --with_box_refine --num_queries 500 --resume crowdhuman_final.pth --epochs 200 --lr_drop 10 and I use the MOT17 as the training set (which is 6.1GB). Are there any changes of parameters or something except --batchsize 2 and --dataset_file mot? I have tried some times about this training progress and always got the same error.

I'd appreciate your help. Thanks!

PeizeSun commented 2 years ago

Hi~ If the batch size can't be reduced, maybe you could try to reduce the image size.