How to change the GPU device? about pointr

[yuxumin]

Use the following command to train a model with GPU:1

bash ./scripts/train.sh 1 --config <config> 

[whuhxb]

@yuxumin OK. Thanks. But when I change into GPU:1, I still met the OOM when running the command: bash ./scripts/train.sh 2 --config ./cfgs/PCN_models/GRNet.yaml --exp_name example

init.py", line 22, in forward
min_x, max_x, min_y, max_y, min_z, max_z, gt_cloud)
RuntimeError: CUDA out of memory. Tried to allocate 512.00 MiB (GPU 0; 31.75 GiB total capacity; 29.93 GiB already allocated; 451.50 MiB free; 29.99 GiB reserved in total by PyTorch)

It seems that the code still run on GPU: 0 ? I'm not sure.

[yuxumin]

It seems that the code still run on GPU: 0 ? I'm not sure.

The code in train.sh is CUDA_VISIBLE_DEVICES=${GPUS} python main.py ${PY_ARGS}. I am sure that the code run on GPU:1

bash ./scripts/train.sh 2 --config ./cfgs/PCN_models/GRNet.yaml --exp_name example

And this will put the model and data on GPU:2

[whuhxb]

@yuxumin OK. I see. I try to run other GPU Cluster nodes, but I met an error like this:

anaconda3/envs/pytorch-PoinTr/lib/python3.7/site-packages/knn_cuda/init.py", line 15, in load_cpp_ext
assert torch.cuda.is_available(), "torch.cuda.is_available() is False."
AssertionError: torch.cuda.is_available() is False.

I use cuda11.1 and pytorch==1.8.0, it seems that the GPU is not available now.

[whuhxb]

@yuxumin Which file to change batchsize of GRNet.yaml and TopNet.yaml?

[yuxumin]

bash ./scripts/train.sh 2 --config ./cfgs/PCN_models/GRNet.yaml --exp_name example

This code means to use the config file in cfgs/PCN_models/GRNet.yaml

[whuhxb]

@yuxumin Is this: total_bs : 32 in cfgs/PCN_models/GRNet.yaml to change the batch size, right?

[yuxumin]

yes

[whuhxb]

@yuxumin I have change it into 2, the OOM still occurs.

[yuxumin]

GRNet requires to calculate gridding loss during the training, which takes a lot of memory of GPU. Can you try to train other models, rather than GRNet

[yuxumin]

Close it since no response. Feel free to re-open it if problems still exist