Make duplicate name test more robust (#762)

* Make duplicate name test more robust

* Resolve ~/.keras issue as well

.travis.yml

@@ -121,6 +121,9 @@ script:
     fi
 
 
+  # prepare ~/.keras folder to avoid race condition
+  - docker exec ${CONTAINER} /bin/sh -c "mkdir -p ~/.keras"
+
   # run unit tests
   - docker exec ${CONTAINER} /bin/sh -c "pip install pytest && cd /horovod/test && (echo test_*.py | xargs -n 1 ${MPIRUN} pytest -v)"
 

test/test_torch.py

@@ -368,7 +368,8 @@ def test_horovod_allreduce_duplicate_name_error(self):
 
         hvd.allreduce_async(tensor, name='duplicate_name')
         try:
-            hvd.allreduce_async(tensor, name='duplicate_name')
+            for i in range(10):
+                hvd.allreduce_async(tensor, name='duplicate_name')
             assert False, 'hvd.allreduce_async did not throw error'
         except (torch.FatalError, ValueError):
             pass
@@ -560,7 +561,8 @@ def test_horovod_allgather_duplicate_name_error(self):
 
         hvd.allgather_async(tensor, name='duplicate_name')
         try:
-            hvd.allgather_async(tensor, name='duplicate_name')
+            for i in range(10):
+                hvd.allgather_async(tensor, name='duplicate_name')
             assert False, 'hvd.allgather_async did not throw error'
         except (torch.FatalError, ValueError):
             pass
@@ -757,7 +759,8 @@ def test_horovod_broadcast_duplicate_name_error(self):
 
         hvd.broadcast_async(tensor, root_rank=0, name='duplicate_name')
         try:
-            hvd.broadcast_async(tensor, root_rank=0, name='duplicate_name')
+            for i in range(10):
+                hvd.broadcast_async(tensor, root_rank=0, name='duplicate_name')
             assert False, 'hvd.broadcast_async did not throw error'
         except (torch.FatalError, ValueError):
             pass