[Distributed] Allow additional horovod DistributedOptimizer args.

Signed-off-by: 泊霆 <[email protected]>

Author: Mesilenceki

tensorflow/python/distribute/hvd_strategy.py

@@ -397,7 +397,15 @@ def __init__(self, learning_rate=0.001, *args, **kwargs):
     else:
         def horovod_optimizer(*args, **kwargs):
             from horovod.tensorflow import DistributedOptimizer
-            return DistributedOptimizer(HvdOptimizer(*args, **kwargs))
+            horovod_args = DistributedOptimizer.__code__.co_varnames
+            horovod_real_kargs = {}
+            candidate_keys = list(kwargs.keys())
+            for kwarg in candidate_keys:
+                if kwarg in horovod_args:
+                    value = kwargs[kwarg]
+                    del kwargs[kwarg]
+                    horovod_real_kargs[kwarg] = value
+            return DistributedOptimizer(HvdOptimizer(*args, **kwargs), **horovod_real_kargs)
         return horovod_optimizer