search

eric-yyjau / pytorch-superpoint

Superpoint Implemented in PyTorch: https://arxiv.org/abs/1712.07629
MIT License
774 stars 164 forks source link

CUDA out of memory on 3090 #89

Open FeiXie8 opened 1 year ago

FeiXie8 commented 1 year ago

when i export lables on coco2014 and set the resolution to 480*640,CUDA out of memory. my gpu is rtx3090,and the memory is 24GB

ghost commented 1 year ago

Hi, @FeiXie8 . Maybe, i think the batchsize is too many when you train the model. Probably, batchsize will be 8 better.

20181313zhang commented 6 months ago

when i export lables on coco2014 and set the resolution to 480*640,CUDA out of memory. my gpu is rtx3090,and the memory is 24GB

hello,are you solve this question?I meet the same problem.

20181313zhang commented 6 months ago

Hi, @FeiXie8 . Maybe, i think the batchsize is too many when you train the model. Probably, batchsize will be 8 better.

where can i change batchsize?

qingyunwudaoletu commented 3 months ago

export.py not batch size reason.

Maybe you need set num_workers = 0, and pin_memory = False

workers_test = training_params.get('workers_test', 0) # 16 test_loader = torch.utils.data.DataLoader( test_set, batch_size=1, shuffle=False,

pin_memory=True,

  num_workers=workers_test,
  #worker_init_fn=worker_init_fn

in loader.py line 138