Multiple GPU training failed

[makecent]

I tried to use 2 GPU for training, but it raised a error:

PYTHONPATH=$PWD:$PYTHONPATH mim train mmaction configs/localization/apn_coralrandom_r3dsony_32x4_10e_thumos14_flow.py --gpus 2 --validate

2021-06-13 16:38:48,309 - mmaction - INFO - workflow: [('train', 1)], max: 10 epochs
Traceback (most recent call last):
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmaction/tools/train.py", line 199, in <module>
    main()
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmaction/tools/train.py", line 195, in main
    meta=meta)
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmaction/apis/train.py", line 163, in train_model
    runner.run(data_loaders, cfg.workflow, cfg.total_epochs, **runner_kwargs)
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/runner/epoch_based_runner.py", line 125, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/runner/epoch_based_runner.py", line 50, in train
    self.run_iter(data_batch, train_mode=True, **kwargs)
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/runner/epoch_based_runner.py", line 30, in run_iter
    **kwargs)
  File "/home/louis/miniconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/parallel/data_parallel.py", line 55, in train_step
    ('MMDataParallel only supports single GPU training, if you need to'
AssertionError: MMDataParallel only supports single GPU training, if you need to train with multiple GPUs, please use MMDistributedDataParallelinstead.

However, I did's manually set the distributed method in my own code. It seems that mim uses the train.py instead of dist_train.sh . How to fix this?

[ZwwWayne]

You did not set the launcher. You should add option --launcher pytorch

[makecent]

You did not set the launcher. You should add option --launcher pytorch

I also found this solution in the doc. But I think this argument is not that clear for demonstrating distributed training.

[ZwwWayne]

dist_train.sh actually set --launcher pytorch by default. Thanks for your advice, we will try to make it simpler in the future.

[makecent]

dist_train.sh actually set --launcher pytorch by default. Thanks for your advice, we will try to make it simpler in the future.

Thanks for your reply. mim is excellent!