Support Mean in DDP Sync (#2568)

* Update converters.py

* Update test_converters.py

* pep8

* pep8 tests

* Update test_datamodules.py

* Update test_converters.py

* Update converters.py

* Update test_datamodules.py

* Update test_converters.py

* Update test_converters.py

* fix tests

* fix ddp tests on windows

* chlog

Co-authored-by: Jirka Borovec <[email protected]>

Author: justusschock

CHANGELOG.md

@@ -27,6 +27,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Added call DataModule hooks implicitly in trainer ([#2755](https://github.com/PyTorchLightning/pytorch-lightning/pull/2755))
 
+- Added support for Mean in DDP Sync ([#2568](https://github.com/PyTorchLightning/pytorch-lightning/pull/2568))
+
 ### Changed
 
 - Truncated long version numbers in progress bar ([#2594](https://github.com/PyTorchLightning/pytorch-lightning/pull/2594))

pytorch_lightning/metrics/converters.py

@@ -234,23 +234,32 @@ def _sync_ddp_if_available(result: Union[torch.Tensor],
         result: the value to sync and reduce (typically tensor or number)
         group: the process group to gather results from. Defaults to all processes (world)
         reduce_op: the reduction operation. Defaults to sum.
+            Can also be a string of 'avg', 'mean' to calculate the mean during reduction.
 
     Return:
         reduced value
     """
 
     if torch.distributed.is_available() and torch.distributed.is_initialized():
+        divide_by_world_size = False
+
         if group is None:
             group = torch.distributed.group.WORLD
 
         if reduce_op is None:
             reduce_op = torch.distributed.ReduceOp.SUM
+        elif isinstance(reduce_op, str) and reduce_op in ('avg', 'mean'):
+            reduce_op = torch.distributed.ReduceOp.SUM
+            divide_by_world_size = True
 
         # sync all processes before reduction
         torch.distributed.barrier(group=group)
         torch.distributed.all_reduce(result, op=reduce_op, group=group,
                                      async_op=False)
 
+        if divide_by_world_size:
+            result = result / torch.distributed.get_world_size(group)
+
     return result
 
 

tests/metrics/test_converters.py

@@ -1,5 +1,6 @@
 import numpy as np
 import pytest
+import sys
 import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
@@ -114,26 +115,44 @@ def _setup_ddp(rank, worldsize):
     dist.init_process_group("gloo", rank=rank, world_size=worldsize)
 
 
-def _ddp_test_fn(rank, worldsize):
+def _ddp_test_fn(rank, worldsize, add_offset: bool, reduction_mean=False):
     _setup_ddp(rank, worldsize)
-    tensor = torch.tensor([1.], device='cuda:0')
-
-    reduced_tensor = _sync_ddp_if_available(tensor)
+    if add_offset:
+        tensor = torch.tensor([float(rank)])
+    else:
+        tensor = torch.tensor([1.], )
+    if reduction_mean:
+        reduced_tensor = _sync_ddp_if_available(tensor, reduce_op='avg')
+
+        manual_reduction = sum([i for i in range(dist.get_world_size())]) / dist.get_world_size()
+        print(reduced_tensor)
+        print(manual_reduction)
+        assert reduced_tensor.item() == manual_reduction
+    else:
+        reduced_tensor = _sync_ddp_if_available(tensor)
 
-    assert reduced_tensor.item() == dist.get_world_size(), \
-        'Sync-Reduce does not work properly with DDP and Tensors'
+        assert reduced_tensor.item() == dist.get_world_size(), \
+            'Sync-Reduce does not work properly with DDP and Tensors'
 
 
-@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="test requires multi-GPU machine")
+@pytest.mark.skipif(sys.platform == "win32" , reason="DDP not available on windows")
 def test_sync_reduce_ddp():
     """Make sure sync-reduce works with DDP"""
     tutils.reset_seed()
     tutils.set_random_master_port()
 
     worldsize = 2
-    mp.spawn(_ddp_test_fn, args=(worldsize,), nprocs=worldsize)
+    mp.spawn(_ddp_test_fn, args=(worldsize, False), nprocs=worldsize)
 
-    # dist.destroy_process_group()
+
+@pytest.mark.skipif(sys.platform == "win32" , reason="DDP not available on windows")
+def test_sync_reduce_ddp_mean():
+    """Make sure sync-reduce works with DDP"""
+    tutils.reset_seed()
+    tutils.set_random_master_port()
+
+    worldsize = 2
+    mp.spawn(_ddp_test_fn, args=(worldsize, True, True), nprocs=worldsize)
 
 
 def test_sync_reduce_simple():
@@ -172,7 +191,7 @@ def _ddp_test_tensor_metric(rank, worldsize):
     _test_tensor_metric(True)
 
 
-@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="test requires multi-GPU machine")
+@pytest.mark.skipif(sys.platform == "win32" , reason="DDP not available on windows")
 def test_tensor_metric_ddp():
     tutils.reset_seed()
     tutils.set_random_master_port()
@@ -212,7 +231,7 @@ def _ddp_test_numpy_metric(rank, worldsize):
     _test_numpy_metric(True)
 
 
-@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="test requires multi-GPU machine")
+@pytest.mark.skipif(sys.platform == "win32" , reason="DDP not available on windows")
 def test_numpy_metric_ddp():
     tutils.reset_seed()
     tutils.set_random_master_port()