diff --git a/ppcls/engine/engine.py b/ppcls/engine/engine.py index 7ab29d8d270d2db14d77fef43d6a273e395c9bea..bc3a2a16036f73a89e954c52d5372c5cef19e39f 100644 --- a/ppcls/engine/engine.py +++ b/ppcls/engine/engine.py @@ -242,6 +242,11 @@ 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