Merge pull request #428 from harshithapv/GrandientDescentOptimizerChange

Fixed a bug that over writes the learning rate when sent as a Tensor.

Author: Oceania2018

src/TensorFlowNET.Core/Train/GradientDescentOptimizer.cs

@@ -35,22 +35,29 @@ public class GradientDescentOptimizer : Optimizer
         /// for changing these values across different invocations of optimizer
         /// functions.
         /// </remarks>
+        private bool _useTensor;
         public GradientDescentOptimizer(float learning_rate, bool use_locking = false, string name = "GradientDescent") 
             : base(learning_rate, use_locking, name)
         {
             _lr = learning_rate;
+            _useTensor = false;
         }
 
         public GradientDescentOptimizer(Tensor learning_rate, bool use_locking = false, string name = "GradientDescent")
             : base(learning_rate, use_locking, name)
         {
             _lr_t = learning_rate;
+            _useTensor = true;
         }
 
         public override void _prepare()
         {
-            var lr = _call_if_callable(_lr);
-            _lr_t = ops.convert_to_tensor(lr, name: "learning_rate");
+            if(!_useTensor)
+            {
+                var lr = _call_if_callable(_lr);
+                _lr_t = ops.convert_to_tensor(lr, name: "learning_rate");
+            }
+            
         }
     }
 }