fix: convert bn to sync_bn

the running_mean and running_var of bn would not be synchronized in dist,
so which leads to bug that eval loss in training is inconsistent with eval only.

ppcls/engine/engine.py

@@ -243,6 +243,11 @@ class Engine(object):
                 level=amp_level,
                 save_dtype='float32')
 
+        # TODO(gaotingquan): convert_sync_batchnorm is not effective
+        # eval loss in training is inconsistent with the eval only if bn is used,
+        # because the running_mean and running_var of bn are not synced in dist.
+        self.model = nn.SyncBatchNorm.convert_sync_batchnorm(self.model)
+
         # for distributed
         world_size = dist.get_world_size()
         self.config["Global"]["distributed"] = world_size != 1