Updated Tensorflow.Net to 0.70.2 with Tensorflow 2.7.0. (#7472)

* Updated Tensorflow.Net to 0.70.2 with Tensorflow 2.7.0.

NumSharp replaced with Tensorflow.NumPy.
TensorShape replaced with Shape, Shape object has dimensions as 64 bit long, check added for casting to 32 bit int alsoTensor constructor using SafeTensorHandle/DangerousGetHandle and TF_DataType not required when casting.

Added StringTensorFactory to wrap addition tensorflow.dll methods required to create Tensors from string based input.

* Common code to ArrayUtils, dotnet added to Sdk install to resolve CI issues.

* Temporarily omit GPU packages

* Removing installation of .NET 9.0

* Update TensorFlow.NET to 0.150.0

* Rollback TensorFlow.NET to 0.110.1 to workaround strong name signing issue

Tensorflow.NET stopped being strong name signed in 0.110.2.  See SciSharp/TensorFlow.NET#1296

* Rollback TensorFlow.Net to 0.100.4

After this version a dependency was introduced on OneOf which is not strong name signed and breaks use on .NETFramework.

* Rollback TensorFlow.Net to 0.100.2

We can't use newer than 0.100.4 due to strong naming issues.

There's a bug in the Session.Dispose introduced by SciSharp/TensorFlow.NET@a7c9a75 (in 0.100.4) which was later fixed in SciSharp/TensorFlow.NET@58de537 but that's after the strong naming regression.

* Use different version for TensorFlow GPU packages

Workaround SciSharp/TensorFlow.NET#1294

* Remove VectorDataViewType overload that accepts long[]

* Update Microsoft.ML.Samples.GPU.csproj

* Workaround build error from TF-GPU package

* Update src/Microsoft.ML.TensorFlow/TensorflowTransform.cs

* Avoid finalizer crash in tests

The TensorFlow.NET session finalizer has a bug where it will crash if run after the finalizer for the `tensorflow` type.

Avoid that by ensuring we dispose all Session objects.

* Fix a couple more missing dispose

---------

Co-authored-by: Ross Halliday <[email protected]>
Co-authored-by: Eric StJohn <[email protected]>

Author: 3 people

Directory.Packages.props

@@ -42,8 +42,8 @@
     <PackageVersion Include="Parquet.Net" Version="$(ParquetDotNetVersion)" />
     <PackageVersion Include="Plotly.NET.CSharp" Version="$(PlotlyNETCSharpVersion)" />
     <PackageVersion Include="SciSharp.TensorFlow.Redist" Version="$(TensorFlowVersion)" />
-    <PackageVersion Include="SciSharp.TensorFlow.Redist-Linux-GPU" Version="$(TensorFlowVersion)" />
-    <PackageVersion Include="SciSharp.TensorFlow.Redist-Windows-GPU" Version="$(TensorFlowVersion)" />
+    <PackageVersion Include="SciSharp.TensorFlow.Redist-Linux-GPU" Version="$(TensorFlowLinuxGpuVersion)" />
+    <PackageVersion Include="SciSharp.TensorFlow.Redist-Windows-GPU" Version="$(TensorFlowWindowsGpuVersion)" />
     <PackageVersion Include="SharpZipLib" Version="$(SharpZipLibVersion)" />
     <PackageVersion Include="SkiaSharp" Version="$(SkiaSharpVersion)" />
     <PackageVersion Include="SkiaSharp.NativeAssets.Linux.NoDependencies" Version="$(SkiaSharpVersion)" />

docs/samples/Microsoft.ML.Samples.GPU/Microsoft.ML.Samples.GPU.csproj

@@ -53,22 +53,27 @@
     <PackageReference Include="SciSharp.TensorFlow.Redist-Linux-GPU" />
   </ItemGroup>
 
-    <ItemGroup>
+  <!-- workaround https://github.com/SciSharp/TensorFlow.NET/issues/1297 -->
+  <Target Name="_EnsureObjFolder" BeforeTargets="FileRestitch">
+    <MakeDir Directories="obj" />
+  </Target>
+  
+  <ItemGroup>
     <Content Include="$(ModelDownloadPath)\ResNet18Onnx\ResNet18.onnx">
       <Link>DnnImageModels\ResNet18Onnx\ResNet18.onnx</Link>
       <CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
     </Content>
   </ItemGroup>
 
-    <ItemGroup>
+ <ItemGroup>
     <Content Include="$(ModelDownloadPath)\ResNetPrepOnnx\ResNetPreprocess.onnx">
       <Link>DnnImageModels\ResNetPrepOnnx\ResNetPreprocess.onnx</Link>
       <CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
     </Content>
   </ItemGroup>
 
-    <ItemGroup>
-      <Folder Include="Dynamic\" />
-    </ItemGroup>
+  <ItemGroup>
+    <Folder Include="Dynamic\" />
+  </ItemGroup>
 
 </Project>

eng/Versions.props

@@ -67,9 +67,13 @@
     <ParquetDotNetVersion>2.1.3</ParquetDotNetVersion>
     <PlotlyNETCSharpVersion>0.11.1</PlotlyNETCSharpVersion>
     <SharpZipLibVersion>1.4.2</SharpZipLibVersion>
-    <TensorflowDotNETVersion>0.20.1</TensorflowDotNETVersion>
+    <!-- TensorFlow.NET cannot be updated past 0.100.2 due to https://github.com/SciSharp/TensorFlow.NET/issues/1296, 
+     and regression https://github.com/SciSharp/TensorFlow.NET/commit/a7c9a75954d219cb606042fcbfbeb1b176781d7e introduced in 0.100.4 -->
+    <TensorflowDotNETVersion>0.100.2</TensorflowDotNETVersion>
     <TensorFlowMajorVersion>2</TensorFlowMajorVersion>
-    <TensorFlowVersion>2.3.1</TensorFlowVersion>
+    <TensorFlowVersion>2.16.0</TensorFlowVersion>
+    <TensorFlowWindowsGpuVersion>2.10.3</TensorFlowWindowsGpuVersion>
+    <TensorFlowLinuxGpuVersion>2.11.1</TensorFlowLinuxGpuVersion>
     <TorchSharpPyBridgeVersion>1.4.1</TorchSharpPyBridgeVersion>
     <AutoGenVersion>0.2.3</AutoGenVersion>
     <SemanticKernelVersion>1.48.0</SemanticKernelVersion>

src/Microsoft.ML.TensorFlow/TensorTypeExtensions.cs

@@ -25,7 +25,7 @@ public static void ToScalar<T>(this Tensor tensor, ref T dst) where T : unmanage
                 return;
             }
 
-            if (typeof(T).as_dtype() != tensor.dtype)
+            if (typeof(T).as_tf_dtype() != tensor.dtype)
                 throw new NotSupportedException();
 
             unsafe
@@ -37,7 +37,7 @@ public static void ToScalar<T>(this Tensor tensor, ref T dst) where T : unmanage
 
         public static void CopyTo<T>(this Tensor tensor, Span<T> values) where T : unmanaged
         {
-            if (typeof(T).as_dtype() != tensor.dtype)
+            if (typeof(T).as_tf_dtype() != tensor.dtype)
                 throw new NotSupportedException();
 
             unsafe

src/Microsoft.ML.TensorFlow/TensorflowTransform.cs

@@ -16,8 +16,8 @@
 using Microsoft.ML.Runtime;
 using Microsoft.ML.TensorFlow;
 using Microsoft.ML.Transforms;
-using NumSharp;
 using Tensorflow;
+using Tensorflow.NumPy;
 using static Microsoft.ML.TensorFlow.TensorFlowUtils;
 using static Tensorflow.Binding;
 using Utils = Microsoft.ML.Internal.Utilities.Utils;
@@ -51,7 +51,7 @@ public sealed class TensorFlowTransformer : RowToRowTransformerBase, IDisposable
         internal readonly DataViewType[] OutputTypes;
         internal readonly TF_DataType[] TFOutputTypes;
         internal readonly TF_DataType[] TFInputTypes;
-        internal readonly TensorShape[] TFInputShapes;
+        internal readonly Shape[] TFInputShapes;
         internal readonly (Operation, int)[] TFInputOperations;
         internal readonly (Operation, int)[] TFOutputOperations;
         internal TF_Output[] TFInputNodes;
@@ -212,14 +212,14 @@ internal TensorFlowTransformer(IHostEnvironment env, TensorFlowEstimator.Options
             env.CheckValue(options, nameof(options));
         }
 
-        private static ITensorValueGetter CreateTensorValueGetter<T>(DataViewRow input, bool isVector, int colIndex, TensorShape tfShape)
+        private static ITensorValueGetter CreateTensorValueGetter<T>(DataViewRow input, bool isVector, int colIndex, Shape tfShape)
         {
             if (isVector)
                 return new TensorValueGetterVec<T>(input, colIndex, tfShape);
             return new TensorValueGetter<T>(input, colIndex, tfShape);
         }
 
-        private static ITensorValueGetter CreateTensorValueGetter(DataViewRow input, TF_DataType tfType, bool isVector, int colIndex, TensorShape tfShape)
+        private static ITensorValueGetter CreateTensorValueGetter(DataViewRow input, TF_DataType tfType, bool isVector, int colIndex, Shape tfShape)
         {
             var type = Tf2MlNetType(tfType);
             return Utils.MarshalInvoke(CreateTensorValueGetter<int>, type.RawType, input, isVector, colIndex, tfShape);
@@ -230,7 +230,7 @@ private static ITensorValueGetter[] GetTensorValueGetters(
             int[] inputColIndices,
             bool[] isInputVector,
             TF_DataType[] tfInputTypes,
-            TensorShape[] tfInputShapes)
+            Shape[] tfInputShapes)
         {
             var srcTensorGetters = new ITensorValueGetter[inputColIndices.Length];
             for (int i = 0; i < inputColIndices.Length; i++)
@@ -331,10 +331,10 @@ private static (Operation, int) GetOperationFromName(string operation, Session s
             return (session.graph.OperationByName(operation), 0);
         }
 
-        internal static (TF_DataType[] tfInputTypes, TensorShape[] tfInputShapes, (Operation, int)[]) GetInputInfo(IHost host, Session session, string[] inputs, int batchSize = 1)
+        internal static (TF_DataType[] tfInputTypes, Shape[] tfInputShapes, (Operation, int)[]) GetInputInfo(IHost host, Session session, string[] inputs, int batchSize = 1)
         {
             var tfInputTypes = new TF_DataType[inputs.Length];
-            var tfInputShapes = new TensorShape[inputs.Length];
+            var tfInputShapes = new Shape[inputs.Length];
             var tfInputOperations = new (Operation, int)[inputs.Length];
 
             int index = 0;
@@ -351,15 +351,15 @@ internal static (TF_DataType[] tfInputTypes, TensorShape[] tfInputShapes, (Opera
                     throw host.ExceptParam(nameof(session), $"Input type '{tfInputType}' of input column '{input}' is not supported in TensorFlow");
 
                 tfInputTypes[index] = tfInputType;
-                tfInputShapes[index] = ((Tensor)inputTensor).TensorShape;
+                tfInputShapes[index] = ((Tensor)inputTensor).shape;
                 tfInputOperations[index] = (inputTensor, inputTensorIndex);
                 index++;
             }
 
             return (tfInputTypes, tfInputShapes, tfInputOperations);
         }
 
-        internal static TensorShape GetTensorShape(TF_Output output, Graph graph, Status status = null)
+        internal static Shape GetTensorShape(TF_Output output, Graph graph, Status status = null)
         {
             if (graph == IntPtr.Zero)
                 throw new ObjectDisposedException(nameof(graph));
@@ -370,12 +370,12 @@ internal static TensorShape GetTensorShape(TF_Output output, Graph graph, Status
             cstatus.Check();
 
             if (n == -1)
-                return new TensorShape(new int[0]);
+                return new Shape(new int[0]);
 
             var dims = new long[n];
             c_api.TF_GraphGetTensorShape(graph, output, dims, dims.Length, cstatus.Handle);
             cstatus.Check();
-            return new TensorShape(dims.Select(x => (int)x).ToArray());
+            return new Shape(dims.Select(x => (int)x).ToArray());
         }
 
         internal static (TF_DataType[] tfOutputTypes, DataViewType[] outputTypes, (Operation, int)[]) GetOutputInfo(IHost host, Session session, string[] outputs, bool treatOutputAsBatched)
@@ -402,7 +402,7 @@ internal static (TF_DataType[] tfOutputTypes, DataViewType[] outputTypes, (Opera
                 // i.e. the first dimension (if unknown) is assumed to be batch dimension.
                 // If there are other dimension that are unknown the transformer will return a variable length vector.
                 // This is the work around in absence of reshape transformer.
-                var idims = shape.dims;
+                var idims = shape.dims.Select(x => checked((int)x)).ToArray();
 
                 int[] dims = idims;
                 if (treatOutputAsBatched)
@@ -517,7 +517,7 @@ public void Dispose()
 
                 if (Session != null && Session != IntPtr.Zero)
                 {
-                    Session.close(); // invoked Dispose()
+                    Session.Dispose();
                 }
             }
             finally
@@ -536,7 +536,7 @@ private sealed class Mapper : MapperBase
             private readonly TensorFlowTransformer _parent;
             private readonly int[] _inputColIndices;
             private readonly bool[] _isInputVector;
-            private readonly TensorShape[] _fullySpecifiedShapes;
+            private readonly Shape[] _fullySpecifiedShapes;
             private readonly ConcurrentBag<Runner> _runners;
 
             public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
@@ -546,7 +546,7 @@ public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
                 _parent = parent;
                 _inputColIndices = new int[_parent.Inputs.Length];
                 _isInputVector = new bool[_parent.Inputs.Length];
-                _fullySpecifiedShapes = new TensorShape[_parent.Inputs.Length];
+                _fullySpecifiedShapes = new Shape[_parent.Inputs.Length];
                 for (int i = 0; i < _parent.Inputs.Length; i++)
                 {
                     if (!inputSchema.TryGetColumnIndex(_parent.Inputs[i], out _inputColIndices[i]))
@@ -570,19 +570,19 @@ public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
                         {
                             vecType = (VectorDataViewType)type;
                             var colTypeDims = vecType.Dimensions.Select(dim => (int)dim).ToArray();
-                            _fullySpecifiedShapes[i] = new TensorShape(colTypeDims);
+                            _fullySpecifiedShapes[i] = new Shape(colTypeDims);
                         }
                         else
                             // for primitive type use default TensorShape
-                            _fullySpecifiedShapes[i] = new TensorShape();
+                            _fullySpecifiedShapes[i] = new Shape(Array.Empty<long>());
                     }
                     else
                     {
                         vecType = (VectorDataViewType)type;
                         var colTypeDims = vecType.Dimensions.Select(dim => (int)dim).ToArray();
                         // If the column is one dimension we make sure that the total size of the TF shape matches.
                         // Compute the total size of the known dimensions of the shape.
-                        int valCount = 1;
+                        long valCount = 1;
                         int numOfUnkDim = 0;
                         foreach (var s in shape)
                         {
@@ -592,7 +592,7 @@ public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
                                 numOfUnkDim++;
                         }
                         // The column length should be divisible by this, so that the other dimensions can be integral.
-                        int typeValueCount = type.GetValueCount();
+                        long typeValueCount = type.GetValueCount();
                         if (typeValueCount % valCount != 0)
                             throw Contracts.Except($"Input shape mismatch: Input '{_parent.Inputs[i]}' has shape {originalShape.ToString()}, but input data is of length {typeValueCount}.");
 
@@ -616,22 +616,22 @@ public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
                             throw Contracts.Except($"Input shape mismatch: Input '{_parent.Inputs[i]}' has shape {originalShape.ToString()}, but input data is of length {typeValueCount}.");
 
                         // Fill in the unknown dimensions.
-                        var l = new int[originalShapeNdim];
+                        var l = new long[originalShapeNdim];
                         for (int ishape = 0; ishape < originalShapeNdim; ishape++)
-                            l[ishape] = originalShapeDims[ishape] == -1 ? (int)d : originalShapeDims[ishape];
-                        _fullySpecifiedShapes[i] = new TensorShape(l);
+                            l[ishape] = originalShapeDims[ishape] == -1 ? (long)d : originalShapeDims[ishape];
+                        _fullySpecifiedShapes[i] = new Shape(l);
                     }
 
                     if (_parent._addBatchDimensionInput)
                     {
                         // ndim of default TensorShape is -1, make originDim to 0 in this case.
                         // after addBatchDimension, input column will be changed: type -> type[]
                         var originDim = _fullySpecifiedShapes[i].ndim < 0 ? 0 : _fullySpecifiedShapes[i].ndim;
-                        var l = new int[originDim + 1];
+                        var l = new long[originDim + 1];
                         l[0] = 1;
                         for (int ishape = 1; ishape < l.Length; ishape++)
                             l[ishape] = _fullySpecifiedShapes[i].dims[ishape - 1];
-                        _fullySpecifiedShapes[i] = new TensorShape(l);
+                        _fullySpecifiedShapes[i] = new Shape(l);
                     }
                 }
 
@@ -720,7 +720,7 @@ private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[
                             UpdateCacheIfNeeded(input.Position, srcTensorGetters, activeOutputColNames, outputCache);
 
                             var tensor = outputCache.Outputs[_parent.Outputs[iinfo]];
-                            var tensorSize = tensor.TensorShape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);
+                            var tensorSize = tensor.shape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);
 
                             var editor = VBufferEditor.Create(ref dst, (int)tensorSize);
                             FetchStringData(tensor, editor.Values);
@@ -735,7 +735,7 @@ private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[
                             UpdateCacheIfNeeded(input.Position, srcTensorGetters, activeOutputColNames, outputCache);
 
                             var tensor = outputCache.Outputs[_parent.Outputs[iinfo]];
-                            var tensorSize = tensor.TensorShape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);
+                            var tensorSize = tensor.shape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);
 
                             var editor = VBufferEditor.Create(ref dst, (int)tensorSize);
 
@@ -821,10 +821,10 @@ private class TensorValueGetter<T> : ITensorValueGetter
         {
             private readonly ValueGetter<T> _srcgetter;
             private readonly T[] _bufferedData;
-            private readonly TensorShape _tfShape;
+            private readonly Shape _tfShape;
             private int _position;
 
-            public TensorValueGetter(DataViewRow input, int colIndex, TensorShape tfShape)
+            public TensorValueGetter(DataViewRow input, int colIndex, Shape tfShape)
             {
                 _srcgetter = input.GetGetter<T>(input.Schema[colIndex]);
                 _tfShape = tfShape;
@@ -864,15 +864,15 @@ public Tensor GetBufferedBatchTensor()
         private class TensorValueGetterVec<T> : ITensorValueGetter
         {
             private readonly ValueGetter<VBuffer<T>> _srcgetter;
-            private readonly TensorShape _tfShape;
+            private readonly Shape _tfShape;
             private VBuffer<T> _vBuffer;
             private T[] _denseData;
             private T[] _bufferedData;
             private int _position;
             private readonly long[] _dims;
             private readonly long _bufferedDataSize;
 
-            public TensorValueGetterVec(DataViewRow input, int colIndex, TensorShape tfShape)
+            public TensorValueGetterVec(DataViewRow input, int colIndex, Shape tfShape)
             {
                 _srcgetter = input.GetGetter<VBuffer<T>>(input.Schema[colIndex]);
                 _tfShape = tfShape;