Engines bn fix (#1310)

* fix * fix * version upadte
catalyst-team · Sep 30, 2021 · 3f18d6b · 3f18d6b
1 parent bd3a0d7
commit 3f18d6b
Show file tree

Hide file tree

Showing 3 changed files with 7 additions and 6 deletions.
diff --git a/catalyst/__version__.py b/catalyst/__version__.py
@@ -1 +1 @@
-__version__ = "21.09rc1"
+__version__ = "21.09"
diff --git a/catalyst/engines/apex.py b/catalyst/engines/apex.py
@@ -370,7 +370,7 @@ class DistributedDataParallelAPEXEngine(DistributedDataParallelEngine):
         address: address to use for backend.
         port: port to use for backend.
         sync_bn: boolean flag for batchnorm synchonization during disributed training.
-            if True, applies PyTorch `convert_sync_batchnorm`_ to the model for native torch
+            if True, applies Apex `convert_syncbn_model`_ to the model for native torch
             distributed only. Default, False.
         ddp_kwargs: parameters for `apex.parallel.DistributedDataParallel`.
             More info here:
@@ -439,9 +439,8 @@ def get_engine(self):
         stages:
             ...
 
-    .. _convert_sync_batchnorm:
-        https://pytorch.org/docs/stable/generated/torch.nn.SyncBatchNorm.html#
-        torch.nn.SyncBatchNorm.convert_sync_batchnorm
+    .. _`convert_syncbn_model`:
+        https://nvidia.github.io/apex/parallel.html#apex.parallel.convert_syncbn_model
     """
 
     def __init__(
@@ -501,7 +500,7 @@ def init_components(
         model = model_fn()
         model = self.sync_device(model)
         if self._sync_bn:
-            model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
+            model = apex.parallel.convert_syncbn_model(model)
 
         criterion = criterion_fn()
         criterion = self.sync_device(criterion)

diff --git a/catalyst/engines/fairscale.py b/catalyst/engines/fairscale.py
@@ -340,6 +340,7 @@ def __init__(
         self,
         address: str = None,
         port: Union[str, int] = None,
+        sync_bn: bool = False,
         ddp_kwargs: Dict[str, Any] = None,
         process_group_kwargs: Dict[str, Any] = None,
         scaler_kwargs: Dict[str, Any] = None,
@@ -348,6 +349,7 @@ def __init__(
         super().__init__(
             address=address,
             port=port,
+            sync_bn=sync_bn,
             ddp_kwargs=ddp_kwargs,
             process_group_kwargs=process_group_kwargs,
         )