why `model.train()` before DDP wrapper?

[toyot-li]

@sshaoshuai Thanks for the awesome toolbox. But I'm not quite clear about why this line here occurs before the DDP model wrapper, even with the simple comment. Could you explain more about this? Thanks. For comparison, in some well-developed 2D detection codebases, DDP training for detectors with frozen parameters seems not to require such a workaround.