module 'torch.distributed' has no attribute 'ReduceOp'

[sctrueew]

Hi everyone,

I installed all requirements and when I run python eval.py I got this error:
module 'torch.distributed' has no attribute 'ReduceOp'
my Torch version is 1.1.0
thanks.

[ycszen]

Could you give more details?
Meanwhile, in the eval.py script, it didn't call the 'torch.distributed' or 'ReduceOp'.
Therefore, I need more details to know why you get this problem.

[jay1009]

I met this problem, too.
But it occurred when I run python train.py.
Can you tell me what happened about it?
thanks.

[ycszen]

@jay1009
You need to use this script to run train.py as follows:

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py

[jay1009]

@ycszen
Thanks for your reply. But I only have one GPU, and after I use
export NGPUS=1
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py
It still met the same problem.
Is it necessary to use 8 gpus to run this program?