test GpuLeakByCNN.

Author: Oceania2018

src/TensorFlowNET.Core/Functions/TapeGradientFunctions.cs

@@ -97,7 +97,8 @@ public void Record(Tensors flat_outputs, Tensors inference_args)
                 .ToArray();
             foreach(var capture in captures_from_forward)
             {
-                _func_graph.Outputs.Add(capture);
+                if (!_func_graph.Outputs.Contains(capture))
+                    _func_graph.Outputs.Add(capture);
             }
 
             var forward_function_name = $"{_FORWARD_PREFIX}_{ops.uid()}";

src/TensorFlowNet.Benchmarks/Leak/GpuLeakByCNN.cs

@@ -6,14 +6,15 @@
 using Tensorflow.Keras;
 using static Tensorflow.Binding;
 using static Tensorflow.KerasApi;
+using BenchmarkDotNet.Attributes;
 
 namespace Tensorflow.Benchmark.Leak
 {
-    class GpuLeakByCNN
+    public class GpuLeakByCNN
     {
         protected static LayersApi layers = new LayersApi();
-
-        public static void Test()
+        [Benchmark]
+        public void Run()
         {
             int num = 50, width = 64, height = 64;
             // if width = 128, height = 128, the exception occurs faster
@@ -29,24 +30,24 @@ public static void Test()
 
             tf.enable_eager_execution();
 
-            var inputss = keras.Input((height, width, 3));
+            var inputs = keras.Input((height, width, 3));
 
-            var inputs = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputss);
-            inputs = layers.MaxPooling2D((2, 2)).Apply(inputs);
+            var layer = layers.Conv2D(32, (3, 3), activation: keras.activations.Relu).Apply(inputs);
+            layer = layers.MaxPooling2D((2, 2)).Apply(layer);
 
-            inputs = layers.Flatten().Apply(inputs);
+            layer = layers.Flatten().Apply(layer);
 
-            var outputs = layers.Dense(10).Apply(inputs);
+            var outputs = layers.Dense(10).Apply(layer);
 
-            var model = keras.Model(inputss, outputs, "gpuleak");
+            var model = keras.Model(inputs, outputs, "gpuleak");
 
             model.summary();
 
             model.compile(loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
              optimizer: keras.optimizers.RMSprop(),
              metrics: new[] { "accuracy" });
 
-            model.fit(inputImages, outLables, epochs: 200);
+            model.fit(inputImages, outLables, batch_size: 1, epochs: 200);
         }
     }
 }

src/TensorFlowNet.Benchmarks/Program.cs

@@ -2,13 +2,16 @@
 using BenchmarkDotNet.Running;
 using System;
 using System.Reflection;
+using Tensorflow.Benchmark.Leak;
 
 namespace TensorFlowBenchmark
 {
     class Program
     {
         static void Main(string[] args)
         {
+            new GpuLeakByCNN().Run();
+
             if (args?.Length > 0)
             {
                 for (int i = 0; i < args.Length; i++)

src/TensorFlowNet.Benchmarks/Tensorflow.Benchmark.csproj

@@ -23,13 +23,19 @@
     <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
   </PropertyGroup>
 
+  <ItemGroup>
+    <Compile Remove="BenchmarkDotNet.Artifacts\**" />
+    <EmbeddedResource Remove="BenchmarkDotNet.Artifacts\**" />
+    <None Remove="BenchmarkDotNet.Artifacts\**" />
+  </ItemGroup>
+
   <ItemGroup>
     <None Remove="tensorflow.dll" />
   </ItemGroup>
 
   <ItemGroup>
     <PackageReference Include="BenchmarkDotNet" Version="0.12.1" />
-    <PackageReference Include="SciSharp.TensorFlow.Redist" Version="2.3.1" />
+    <PackageReference Include="SciSharp.TensorFlow.Redist-Windows-GPU" Version="2.3.1" />
   </ItemGroup>
 
   <ItemGroup>