[Feature]: support CPU training with MMDataParallel (#972)

* support for CPU training

* Update .pre-commit-config.yaml

* Update data_parallel.py

mmcv/parallel/data_parallel.py

@@ -49,7 +49,7 @@ def train_step(self, *inputs, **kwargs):
             # We add the following line thus the module could gather and
             # convert data containers as those in GPU inference
             inputs, kwargs = self.scatter(inputs, kwargs, [-1])
-            return self.module.train_step(*inputs, **kwargs)
+            return self.module.train_step(*inputs[0], **kwargs[0])
 
         assert len(self.device_ids) == 1, \
             ('MMDataParallel only supports single GPU training, if you need to'
@@ -71,7 +71,7 @@ def val_step(self, *inputs, **kwargs):
             # We add the following line thus the module could gather and
             # convert data containers as those in GPU inference
             inputs, kwargs = self.scatter(inputs, kwargs, [-1])
-            return self.module.val_step(*inputs, **kwargs)
+            return self.module.val_step(*inputs[0], **kwargs[0])
 
         assert len(self.device_ids) == 1, \
             ('MMDataParallel only supports single GPU training, if you need to'