how to set up gpu-id when training with single gpu based on version 1.1

[JiSuanJiDaWang]

Hi, I used the new version of mmdetection3d(1.1). When training with a single GPU, I could not find the gpu-id setting in the tools/train.py file. Please give me some information on how to set it up. Thanks.

[ZCMax]

You can try just use CUDA_VISIBLE_DEVICES=1 python train.py xxxx

[JiSuanJiDaWang]

Thanks! I also noticed in the configs/base/datset file, the new version variable 'samples_per_gpu' and 'workers_per_gpu'. I wonder what should I do if I want to set up the variable. Just simply change the batch_size and num_workers?

[ZCMax]

I do not understand what is your meaning about 'set up the variable'

[JiSuanJiDaWang]

I mean if I want to change the number of samples on each GPU or the number of GPU going to be used. Should I just modify batch_size and the num_workers ?

[ZCMax]

'samples_per_gpu' represents the number of samples on each GPU, but if you want to change the number of GPU going to be used, you can use CUDA_VISIBLE_DEVICES=1, 2, 3, 4... to change the gpu the program can "see", workers_per_gpu does not mean the the number of GPUs going to be used.

[JiSuanJiDaWang]

got it. Thanks a lot!