TF.NET backend in progress

SciSharp · Apr 16, 2019 · bf470be · bf470be
1 parent 029d84f
commit bf470be
Show file tree

Hide file tree

Showing 13 changed files with 162 additions and 61 deletions.
diff --git a/Backends/SiaNet.Backend.TensorFlow/SiaNet.Backend.TensorFlow.csproj b/Backends/SiaNet.Backend.TensorFlow/SiaNet.Backend.TensorFlow.csproj
@@ -23,7 +23,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="TensorFlow.NET" Version="0.5.2" />
+    <PackageReference Include="TensorFlow.NET" Version="0.6.0" />
   </ItemGroup>
 
   <ItemGroup>

diff --git a/Backends/SiaNet.Backend.TensorFlow/SiaNetBackend.cs b/Backends/SiaNet.Backend.TensorFlow/SiaNetBackend.cs
@@ -123,7 +123,7 @@ public Engine.DataType GetDataType(SiaTensor x)
 
         public SiaTensor CreateVariable(float[] data, long[] shape, string name = "")
         {
-            return null;
+            return Out(tf.Variable<Array>(data).value());
         }
 
         public SiaTensor Reshape(SiaTensor x, params long[] shape)
@@ -348,7 +348,7 @@ public SiaTensor Round(SiaTensor x)
 
         public SiaTensor Sin(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.sin(In(x)));
         }
 
         public SiaTensor Cos(SiaTensor x)
@@ -358,37 +358,37 @@ public SiaTensor Cos(SiaTensor x)
 
         public SiaTensor Tan(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.tan(In(x)));
         }
 
         public SiaTensor Asin(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.asin(In(x)));
         }
 
         public SiaTensor Acos(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.acos(In(x)));
         }
 
         public SiaTensor Atan(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.atan(In(x)));
         }
 
         public SiaTensor Sinh(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.sinh(In(x)));
         }
 
         public SiaTensor Cosh(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.cosh(In(x)));
         }
 
         public SiaTensor Tanh(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.tanh(In(x)));
         }
 
         public SiaTensor Sigmoid(SiaTensor x)
@@ -398,47 +398,60 @@ public SiaTensor Sigmoid(SiaTensor x)
 
         public SiaTensor Pow(SiaTensor x, float value)
         {
-            throw new NotImplementedException();
+            return Out(tf.pow(In(x), In(value, x.Shape)));
         }
 
         public SiaTensor Square(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.pow(In(x), 2));
         }
 
         public SiaTensor Clip(SiaTensor x, float min, float max)
         {
-            throw new NotImplementedException();
+            return Out(tf._clip_by_value(In(x), In(min, x.Shape), In(max, x.Shape)));
         }
 
         public float Sum(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return tf.reduce_sum(In(x));
         }
 
         public SiaTensor Sum(SiaTensor x, int dim)
         {
-            throw new NotImplementedException();
+            dim = dim < 0 ? x.DimCount + dim : dim;
+            return Out(tf.sum(In(x), dim, true));
         }
 
-        public SiaTensor Sum(SiaTensor x, params int[] dim)
+        public SiaTensor Sum(SiaTensor x, params int[] dims)
         {
-            throw new NotImplementedException();
+            foreach (var item in dims)
+            {
+                int dim = item < 0 ? x.DimCount + item : item;
+                x = Sum(x, item);
+            }
+
+            return x;
         }
 
         public float Max(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return math_ops.reduce_max(In(x));
         }
 
         public SiaTensor Max(SiaTensor x, int dim)
         {
-            throw new NotImplementedException();
+            dim = dim < 0 ? x.DimCount + dim : dim;
+            return Out(math_ops.reduce_max(In(x), new int[] { dim }, true));
         }
 
-        public SiaTensor Max(SiaTensor x, params int[] dim)
+        public SiaTensor Max(SiaTensor x, params int[] dims)
         {
-            throw new NotImplementedException();
+            for (int i = 0; i < dims.Length; i++)
+            {
+                dims[i] = dims[i] < 0 ? x.DimCount + dims[i] : dims[i];
+            }
+
+            return Out(math_ops.reduce_max(In(x), dims, true));
         }
 
         public float Min(SiaTensor x)
@@ -458,62 +471,70 @@ public SiaTensor Min(SiaTensor x, params int[] dim)
 
         public float Mean(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return math_ops.reduce_mean(In(x));
         }
 
         public SiaTensor Mean(SiaTensor x, int dim)
         {
-            throw new NotImplementedException();
+            dim = dim < 0 ? x.DimCount + dim : dim;
+            return Out(math_ops.reduce_mean(In(x), new int[] { dim }, true));
         }
 
-        public SiaTensor Mean(SiaTensor x, params int[] dim)
+        public SiaTensor Mean(SiaTensor x, params int[] dims)
         {
-            throw new NotImplementedException();
+            for (int i = 0; i < dims.Length; i++)
+            {
+                dims[i] = dims[i] < 0 ? x.DimCount + dims[i] : dims[i];
+            }
+
+            return Out(math_ops.reduce_mean(In(x), dims, true));
         }
 
         public SiaTensor Argmax(SiaTensor x, int dim = 0)
         {
-            throw new NotImplementedException();
+            dim = dim < 0 ? x.DimCount + dim : dim;
+            return Out(tf.arg_max(In(x), dim));
         }
 
         public SiaTensor Argmin(SiaTensor x, int dim = 0)
         {
-            throw new NotImplementedException();
+            dim = dim < 0 ? x.DimCount + dim : dim;
+            return Out(tf.arg_min(In(x), dim));
         }
 
         public SiaTensor Maximum(SiaTensor a, SiaTensor b)
         {
-            throw new NotImplementedException();
+            return Out(tf.maximum(In(a), In(b)));
         }
 
         public SiaTensor Maximum(SiaTensor a, float b)
         {
-            throw new NotImplementedException();
+            return Out(tf.maximum(In(a), In(b, a.Shape)));
         }
 
         public SiaTensor Minimum(SiaTensor a, SiaTensor b)
         {
-            throw new NotImplementedException();
+            return Out(tf.minimum(In(a), In(b)));
         }
 
         public SiaTensor Minimum(SiaTensor a, float b)
         {
-            throw new NotImplementedException();
+            return Out(tf.minimum(In(a), In(b, a.Shape)));
         }
 
         public SiaTensor Transpose(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return Out(tf.transpose(In(x), new int[] { 1, 0 }));
         }
 
         public SiaTensor Transpose(SiaTensor x, params int[] dims)
         {
-            throw new NotImplementedException();
+            return Out(tf.transpose(In(x), dims));
         }
 
         public SiaTensor Dot(SiaTensor a, SiaTensor b)
         {
-            throw new NotImplementedException();
+            return Out(tf.matmul(In(a), In(b)));
         }
 
         public SiaTensor Diag(SiaTensor x)
@@ -523,17 +544,18 @@ public SiaTensor Diag(SiaTensor x)
 
         public SiaTensor Softmax(SiaTensor x, int axis = -1)
         {
-            throw new NotImplementedException();
+            return Out(tf.nn.softmax(In(x), axis));
         }
 
         public SiaTensor Softplus(SiaTensor x, int axis = -1)
         {
-            throw new NotImplementedException();
+            return Log((Exp(x) + 1));
         }
 
         public SiaTensor L2Normalize(SiaTensor x, int axis = -1)
         {
-            throw new NotImplementedException();
+            var y = Max(Sum(Square(x), axis), axis);
+            return x / Sqrt(y);
         }
 
         public SiaTensor Im2Col(SiaTensor x, Tuple<int, int> kernalSize, int padding = 1, int stride = 1)
@@ -558,17 +580,17 @@ public SiaTensor SliceCols(SiaTensor x, long start, long end)
 
         public Array GetArray(SiaTensor x)
         {
-            throw new NotImplementedException();
+            return In(x).Data<float>();
         }
 
         public void Dispose(SiaTensor x)
         {
-            throw new NotImplementedException();
+            In(x).Dispose();
         }
 
         public ActivationFunc GetActFunc()
         {
-            throw new NotImplementedException();
+            return new SiaNetActivations(this);
         }
     }
 }
diff --git a/Backends/SiaNet.Backend.Torch/SiaNet.Backend.Torch.csproj b/Backends/SiaNet.Backend.Torch/SiaNet.Backend.Torch.csproj
@@ -8,6 +8,10 @@
     <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
   </PropertyGroup>
 
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|AnyCPU'">
+    <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+  </PropertyGroup>
+
   <ItemGroup>
     <None Include="TorchSharp.csproj" />
   </ItemGroup>

diff --git a/Examples/BackendTester/Program.cs b/Examples/BackendTester/Program.cs
@@ -1,8 +1,10 @@
-using NumSharp.Core;
+using CNTK;
+using NumSharp;
 using SiaNet;
 using SiaNet.Engine;
 using SiaNet.Initializers;
 using System;
+using System.Collections.Generic;
 using System.Linq;
 
 namespace BackendTester
@@ -11,6 +13,15 @@ class Program
     {
         static void Main(string[] args)
         {
+            int[] shape = new int[] { 6, 3 };
+            float[] data = new float[] { 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 5, 6, 7, 8, 9 };
+
+            NDArrayView array = new NDArrayView(shape, data, DeviceDescriptor.CPUDevice);
+            Variable variable = new Variable(shape, VariableKind.Parameter, CNTK.DataType.Float, array, false, new AxisVector(), false, "", "");
+            var slicedData = CNTKLib.Slice(variable, AxisVector.Repeat(new Axis(0), 1), IntVector.Repeat(1, 1), IntVector.Repeat(3, 1));
+            var resultArray = GetArray(slicedData);
+
+
             Global.UseEngine(SiaNet.Backend.CNTKLib.SiaNetBackend.Instance, DeviceType.CPU);
             var K = Global.CurrentBackend;
 
@@ -25,5 +36,30 @@ static void Main(string[] args)
             Console.ReadLine();
 
         }
+
+        public static Array GetArray(Variable xvar)
+        {
+            Value v = null;
+            if (xvar.IsOutput)
+            {
+                var f = xvar.ToFunction();
+
+                var plist = f.Parameters();
+                Dictionary<Variable, Value> inputs = new Dictionary<Variable, Value>();
+                Dictionary<Variable, Value> outputs = new Dictionary<Variable, Value>()
+                {
+                    { f, null}
+                };
+
+                f.Evaluate(inputs, outputs, DeviceDescriptor.CPUDevice);
+                v = outputs.FirstOrDefault().Value;
+            }
+            else
+            {
+                v = new Value(xvar.GetValue());
+            }
+
+            return v.GetDenseData<float>(xvar)[0].ToArray();
+        }
     }
 }
diff --git a/Examples/BasicClassificationWithTitanicDataset/Program.cs b/Examples/BasicClassificationWithTitanicDataset/Program.cs
@@ -16,11 +16,11 @@ class Program
         static void Main(string[] args)
         {
             //Setup Engine
-            Global.UseEngine(SiaNet.Backend.MxNetLib.SiaNetBackend.Instance, DeviceType.Default);
+            Global.UseEngine(SiaNet.Backend.ArrayFire.SiaNetBackend.Instance, DeviceType.Default);
 
             var train = LoadTrain(); //Load train data
             var test = LoadTest(); //Load test data
-
+            
             var model = new Sequential();
             model.EpochEnd += Model_EpochEnd;
             model.Add(new Dense(128, ActType.ReLU));
@@ -31,10 +31,9 @@ static void Main(string[] args)
             model.Compile(OptimizerType.Adam, LossType.BinaryCrossEntropy, MetricType.BinaryAccurary);
 
             // Perform training with train and val dataset
-            model.Train(train, epochs: 100, batchSize: 32);
+            model.Train(train, epochs: 100, batchSize: 200);
 
-            //var prediction = model.Predict(test);
-            //TOps.Round(prediction).Print();
+            var prediction = model.Predict(test);
         }
 
         private static void Model_EpochEnd(object sender, EpochEndEventArgs e)

diff --git a/Examples/BostonHousingRegressionExample/Program.cs b/Examples/BostonHousingRegressionExample/Program.cs
@@ -13,7 +13,7 @@ class Program
         static void Main(string[] args)
         {
             //Setup Engine
-            Global.UseEngine(SiaNet.Backend.MxNetLib.SiaNetBackend.Instance, DeviceType.CPU);
+            Global.UseEngine(SiaNet.Backend.ArrayFire.SiaNetBackend.Instance, DeviceType.CPU);
 
             //Load Train and Test CSV data
             var ds = LoadTrain("./train.csv");

diff --git a/Examples/GettingStarted/Program.cs b/Examples/GettingStarted/Program.cs
@@ -38,7 +38,7 @@ static void Main(string[] args)
             predX.Load(0, 0, 0, 1); //Result should be 0 and 1
 
             var rawPred = model.Predict(predX);
-            Global.CurrentBackend.Round(rawPred).Print();
+
             Console.ReadLine();
         }