AttributeError: module 'torch.distributed' has no attribute 'is_initialized'

🐛 Bug

When running the webcam demo on macOS 10.14.1 as so:

python webcam.py --min-image-size 300 MODEL.DEVICE cpu

I get the error:

AttributeError: module 'torch.distributed' has no attribute 'is_initialized'

To Reproduce

Steps to reproduce the behavior:

1. Followed install guide here: https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/INSTALL.md
2. Use the following command:

python webcam.py --min-image-size 300 MODEL.DEVICE cpu

Expected behavior

I expect this to run the webcam demo.

Environment

• PyTorch Version (e.g., 1.0): 1.0 (nightly)
• OS (e.g., Linux): macOS 10.14.1
• How you installed PyTorch (conda, pip, source): Conda
• Build command you used (if compiling from source):
• Python version: 3.7
• CUDA/cuDNN version: N/A
• GPU models and configuration: N/A
• Any other relevant information:

Additional context

I fixed this issue by changing all instances of is_initialized() to is_available() in maskrcnn-benchmark/maskrcnn_benchmark/utils/comm.py. I’m not sure if this is what is intended or if I am installing something incorrectly. After making that change, I was able to run the webcam demo.

[congve1]

I’ve got the same error. You could check what torch.distributed.is_available() returns. If False, it’s obvious that your pytorch version doesn’t support distributing training, and the code may face some error

[fmassa]

@apacha good catch! Could you send a PR fixing it?

Maybe a better solution would be to use https://github.com/facebookresearch/maskrcnn-benchmark/blob/1818bb2f457c082dfd3759e6331e40f97d3059c7/maskrcnn_benchmark/utils/comm.py#L13-L18