VBLC

VBLC copied to clipboard

Hi, Thank you for sharing the code of your work.

While reviewing the './tools/train.py' script, I noticed that the multi-GPU mode is not supported.

I was wondering if there is an alternative way for me to train the code using MMDistributedDataParallel. I have NVIDIA TITAN V (12GB) GPUs, which cannot train the model based on Transformer in single GPU.

 if args.gpus is not None:
        cfg.gpu_ids = range(4)    
        warnings.warn('`--gpus` is deprecated because we only support '
                      'single GPU mode in non-distributed training. '
                      'Use `gpus=1` now.')
    if args.gpu_ids is not None:
        cfg.gpu_ids = args.gpu_ids[0:3]
        warnings.warn('`--gpu-ids` is deprecated, please use `--gpu-id`. '
                      'Because we only support single GPU mode in '
                      'non-distributed training. Use the first GPU '
                      'in `gpu_ids` now.')

Thank you for response in advance.

Hi gyuwonchoi, Thanks for your interest in our work!

We haven't tried training on multiple GPUs, but I assume a simple answer is yes.

We base our method on MMSegmentation, and here is a documentation from it on how to train on multiple GPUs.

Since we use run_experiments.py in place of tools/train.py, there should be some difference in using, e.g., the entry point in tools/dist_train.sh should change accordingly.

Also, modification might be made to samples_per_gpu (e.g., from 2 to 1 for two GPUs) to keep the training batch size(2 source + 2 target = 4 in total). Explanations could be found here.

Apologies for not having time to delve into this now. Any feedbacks are welcome if you are willing to try it out!

Best.