[bug] error when using python tools/train.py to train ReDet in MMRotate.

[zytx121]

when the self.distributed equals to False, revert_sync_batchnorm function cannot work normally.

mmengine/mmengine/runner/runner.py

Line 848 in bda92b4

model = revert_sync_batchnorm(model)

We argue this is caused by ReDet replacing torch.nn with e2cnn.nn.

https://github.com/open-mmlab/mmrotate/blob/dev-1.x/mmrotate/models/backbones/re_resnet.py

[HAOCHENYE]

Can you provide a more detailed description of the error? Does revert_sync_batchnorm raise an error or not convert the SyncBN to BN?

[zytx121]

Yes, revert_sync_batchnorm raises an error when converting the SyncBN to BN. However, there is no SyncBatchNorm module in e2cnn. The error occurs by the add_module function. In e2cnn, R2Conv is the smallest unit, which cannot be split. When we try to recreate R2Conv by add_module, this line raise an error. It looks like R2Conv don't support being created by his child modules.

Maybe, we need to allow the user to close revert_sync_batchnorm in the runner.

[HAOCHENYE]

It makes sense to me. Will it be more proper to fallback not revert SyncBN if revert_sync_batchnorm failed?

[zytx121]

It makes sense to me. Will it be more proper to fallback not revert SyncBN if revert_sync_batchnorm failed?

That sounds like a nice idea.