fix object reference issue for _AggregatedGrads #303

Author: Oceania2018

src/TensorFlowHub/TensorFlowHub.csproj

@@ -1,4 +1,4 @@
-<Project Sdk="Microsoft.NET.Sdk">
+<Project Sdk="Microsoft.NET.Sdk">
   <PropertyGroup>
     <AssemblyName>TensorFlow.Net.Hub</AssemblyName>
     <RootNamespace>Tensorflow.Hub</RootNamespace>
@@ -8,7 +8,7 @@
     <ProjectReference Include="..\..\src\TensorFlowNET.Core\TensorFlowNET.Core.csproj" />
   </ItemGroup>
   <ItemGroup>
-    <PackageReference Include="NumSharp" Version="0.10.4" />
+    <PackageReference Include="NumSharp" Version="0.10.5" />
     <PackageReference Include="sharpcompress" Version="0.23.0" />
   </ItemGroup>
 </Project>

src/TensorFlowNET.Core/Gradients/gradients_util.cs

@@ -137,7 +137,7 @@ public static Tensor[] _GradientsHelper(Tensor[] ys,
                                         if (loop_state != null)
                                             ;
                                         else
-                                            out_grads[i] = control_flow_ops.ZerosLikeOutsideLoop(op, i);
+                                            out_grads[i] = new List<Tensor> { control_flow_ops.ZerosLikeOutsideLoop(op, i) };
                                     }
                                 }
 
@@ -146,7 +146,7 @@ public static Tensor[] _GradientsHelper(Tensor[] ys,
                                     string name1 = scope1;
                                     if (grad_fn != null)
                                     {
-                                        in_grads = _MaybeCompile(grad_scope, op, out_grads, null, grad_fn);
+                                        in_grads = _MaybeCompile(grad_scope, op, out_grads[0].ToArray(), null, grad_fn);
                                         _VerifyGeneratedGradients(in_grads, op);
                                     }
 
@@ -310,10 +310,9 @@ private static IEnumerable<Tensor> _NonEagerInputs(Operation op, Tensor[] xs)
                 yield return op.inputs[i];
         }
 
-        private static Tensor[] _AggregatedGrads(Dictionary<string, List<List<Tensor>>> grads, Operation op, string gradient_uid, object loop_state, int aggregation_method = 0)
+        private static List<List<Tensor>> _AggregatedGrads(Dictionary<string, List<List<Tensor>>> grads, Operation op, string gradient_uid, object loop_state, int aggregation_method = 0)
         {
             var out_grads = _GetGrads(grads, op);
-            var return_grads = new Tensor[out_grads.Count];
 
             foreach (var (i, out_grad) in enumerate(out_grads))
             {
@@ -334,21 +333,21 @@ private static Tensor[] _AggregatedGrads(Dictionary<string, List<List<Tensor>>>
                             throw new ValueError("_AggregatedGrads out_grad.Length == 0");
                         }
 
-                        return_grads[i] = out_grad[0];
+                        out_grads[i] = out_grad;
                     }
                     else
                     {
                         used = "add_n";
-                        return_grads[i] = _MultiDeviceAddN(out_grad.ToArray(), gradient_uid);
+                        out_grads[i] = new List<Tensor> { _MultiDeviceAddN(out_grad.ToArray(), gradient_uid) };
                     }
                 }
                 else
                 {
-                    return_grads[i] = null;
+                    out_grads[i] = null;
                 }
             }
 
-            return return_grads;
+            return out_grads;
         }
 
         /// <summary>
@@ -362,18 +361,18 @@ private static Tensor _MultiDeviceAddN(Tensor[] tensor_list, string gradient_uid
             // Basic function structure comes from control_flow_ops.group().
             // Sort tensors according to their devices.
             var tensors_on_device = new Dictionary<string, List<Tensor>>();
-            
+
             foreach (var tensor in tensor_list)
             {
                 if (!tensors_on_device.ContainsKey(tensor.Device))
                     tensors_on_device[tensor.Device] = new List<Tensor>();
 
                 tensors_on_device[tensor.Device].Add(tensor);
             }
-                
+
             // For each device, add the tensors on that device first.
             var summands = new List<Tensor>();
-            foreach(var dev in tensors_on_device.Keys)
+            foreach (var dev in tensors_on_device.Keys)
             {
                 var tensors = tensors_on_device[dev];
                 ops._colocate_with_for_gradient(tensors[0].op, gradient_uid, ignore_existing: true);

test/TensorFlowNET.Examples/ImageProcessing/DigitRecognitionRNN.cs

@@ -28,7 +28,7 @@ namespace TensorFlowNET.Examples.ImageProcess
     /// </summary>
     public class DigitRecognitionRNN : IExample
     {
-        public bool Enabled { get; set; } = true;
+        public bool Enabled { get; set; } = false;
         public bool IsImportingGraph { get; set; } = false;
 
         public string Name => "MNIST RNN";

test/TensorFlowNET.UnitTest/GradientTest.cs

@@ -2,6 +2,7 @@
 using NumSharp;
 using System.Linq;
 using Tensorflow;
+using static Tensorflow.Python;
 
 namespace TensorFlowNET.UnitTest
 {
@@ -28,6 +29,38 @@ public void Gradients()
             Assert.AreEqual(g[1].name, "gradients/Fill:0");
         }
 
+        [TestMethod]
+        public void Gradient2x()
+        {
+            var graph = tf.Graph().as_default();
+            with(tf.Session(graph), sess => {
+                var x = tf.constant(7.0f);
+                var y = x * x * tf.constant(0.1f);
+
+                var grad = tf.gradients(y, x);
+                Assert.AreEqual(grad[0].name, "gradients/AddN:0");
+
+                float r = sess.run(grad[0]);
+                Assert.AreEqual(r, 1.4f);
+            });
+        }
+
+        [TestMethod]
+        public void Gradient3x()
+        {
+            var graph = tf.Graph().as_default();
+            with(tf.Session(graph), sess => {
+                var x = tf.constant(7.0f);
+                var y = x * x * x * tf.constant(0.1f);
+
+                var grad = tf.gradients(y, x);
+                Assert.AreEqual(grad[0].name, "gradients/AddN:0");
+
+                float r = sess.run(grad[0]);
+                Assert.AreEqual(r, 14.700001f);
+            });
+        }
+
         [TestMethod]
         public void StridedSlice()
         {