Add Preprocessing.timeseries_dataset_from_array

src/TensorFlowNET.Core/Binding.Util.cs

@@ -155,6 +155,8 @@ namespace Tensorflow
         {
             switch (a)
             {
+                case Tensor tensor:
+                    return tensor.shape[0];
                 case Tensors arr:
                     return arr.Length;
                 case Array arr:

src/TensorFlowNET.Core/Tensorflow.Binding.csproj

@@ -5,7 +5,7 @@
     <AssemblyName>TensorFlow.NET</AssemblyName>
     <RootNamespace>Tensorflow</RootNamespace>
     <TargetTensorFlow>2.2.0</TargetTensorFlow>
-    <Version>0.40.1</Version>
+    <Version>0.50.0</Version>
     <LangVersion>8.0</LangVersion>
     <Authors>Haiping Chen, Meinrad Recheis, Eli Belash</Authors>
     <Company>SciSharp STACK</Company>
@@ -19,7 +19,7 @@
     <Description>Google's TensorFlow full binding in .NET Standard.
 Building, training and infering deep learning models.
 https://tensorflownet.readthedocs.io</Description>
-    <AssemblyVersion>0.40.1.0</AssemblyVersion>
+    <AssemblyVersion>0.50.0.0</AssemblyVersion>
     <PackageReleaseNotes>tf.net 0.20.x and above are based on tensorflow native 2.x.
 
 * Eager Mode is added finally.
@@ -31,8 +31,9 @@ https://tensorflownet.readthedocs.io</Description>
 TensorFlow .NET v0.3x is focused on making more Keras API works.
 Keras API is a separate package released as TensorFlow.Keras.
 
-tf.net 0.4x.x aligns with TensorFlow v2.4.1 native library.</PackageReleaseNotes>
-    <FileVersion>0.40.1.0</FileVersion>
+tf.net 0.4x.x aligns with TensorFlow v2.4.1 native library.
+tf.net 0.5x.x aligns with TensorFlow v2.5.x native library.</PackageReleaseNotes>
+    <FileVersion>0.50.0.0</FileVersion>
     <PackageLicenseFile>LICENSE</PackageLicenseFile>
     <PackageRequireLicenseAcceptance>true</PackageRequireLicenseAcceptance>
     <SignAssembly>true</SignAssembly>

src/TensorFlowNET.Core/Tensors/Tensor.Index.cs

@@ -120,6 +120,35 @@ namespace Tensorflow
             });
         }
 
+        public Tensor this[params Tensor[] slice]
+        {
+            get
+            {
+                var args = tensor_util.ParseSlices(slice);
+
+                return tf_with(ops.name_scope(null, "strided_slice", args), scope =>
+                {
+                    string name = scope;
+
+                    var tensor = gen_array_ops.strided_slice(
+                        this,
+                        args.PackedBegin,
+                        args.PackedEnd,
+                        args.PackedStrides,
+                        begin_mask: args.BeginMask,
+                        end_mask: args.EndMask,
+                        shrink_axis_mask: args.ShrinkAxisMask,
+                        new_axis_mask: args.NewAxisMask,
+                        ellipsis_mask: args.EllipsisMask,
+                        name: name);
+
+                    tensor.OriginalVarSlice = args;
+
+                    return tensor;
+                });
+            }
+        }
+
         public Tensor slice(int start)
         {
             var slice_spec = new int[] { start };

src/TensorFlowNET.Core/Tensors/constant_op.cs

@@ -146,9 +146,7 @@ namespace Tensorflow
             }
 
             if (dtype == TF_DataType.TF_STRING && value is byte[] bytes)
-            {
                 return new EagerTensor(bytes, ctx.DeviceName, TF_DataType.TF_STRING);
-            }
 
             switch (value)
             {

src/TensorFlowNET.Core/Tensors/tensor_util.cs

@@ -673,5 +673,38 @@ would not be rank 1.", tensor.op.get_attr("axis")));
                 NewAxisMask = new_axis_mask
             };
         }
+
+        public static ParsedSliceArgs ParseSlices(Tensor[] slices)
+        {
+            var begin = new List<Tensor>();
+            var end = new List<Tensor>();
+            var strides = new List<Tensor>();
+
+            var index = 0;
+            var (new_axis_mask, shrink_axis_mask) = (0, 0);
+            var (begin_mask, end_mask) = (0, 0);
+            var ellipsis_mask = 0;
+
+            foreach (var s in slices)
+            {
+                begin.Add(s);
+                end.Add(s + 1);
+                shrink_axis_mask |= (1 << index);
+                strides.Add(tf.constant(1, dtype: s.dtype));
+                index += 1;
+            }
+
+            return new ParsedSliceArgs
+            {
+                PackedBegin = array_ops.stack(begin),
+                PackedEnd = array_ops.stack(end),
+                PackedStrides = array_ops.stack(strides),
+                BeginMask = begin_mask,
+                EndMask = end_mask,
+                EllipsisMask = ellipsis_mask,
+                ShrinkAxisMask = shrink_axis_mask,
+                NewAxisMask = new_axis_mask
+            };
+        }
     }
 }

src/TensorFlowNET.Keras/Preprocessings/Preprocessing.image_dataset_from_directory.cs

@@ -1,4 +1,6 @@
 using static Tensorflow.KerasApi;
+using static Tensorflow.Binding;
+using NumSharp; 
 
 namespace Tensorflow.Keras
 {
@@ -41,7 +43,7 @@ namespace Tensorflow.Keras
             int num_channels = 0;
             if (color_mode == "rgb")
                 num_channels = 3;
-            
+
             var (image_paths, label_list, class_name_list) = keras.preprocessing.dataset_utils.index_directory(directory,
                 labels,
                 formats: WHITELIST_FORMATS,
@@ -90,5 +92,50 @@ namespace Tensorflow.Keras
             dataset.class_names = class_name_list;
             return dataset;
         }
+
+        /// <summary>
+        /// Creates a dataset of sliding windows over a timeseries provided as array.
+        /// </summary>
+        /// <param name="data"></param>
+        /// <param name="sequence_length"></param>
+        /// <param name="sequence_stride"></param>
+        /// <param name="sampling_rate"></param>
+        /// <param name="batch_size"></param>
+        /// <param name="shuffle"></param>
+        /// <returns></returns>
+        public IDatasetV2 timeseries_dataset_from_array(Tensor data, int sequence_length,
+            int sequence_stride = 1, int sampling_rate = 1, int batch_size = 128,
+            bool shuffle = false, int seed = (int)1e6, int start_index = 0, int? end_index = null)
+        {
+            if (!end_index.HasValue)
+                end_index = len(data);
+
+            var num_seqs = end_index.Value - start_index - (sequence_length * sampling_rate) + 1;
+            var index_dtype = num_seqs < 2147483647 ? tf.int32 : tf.int64;
+            var start_positions = np.arange(0, num_seqs, sequence_stride);
+            if (shuffle)
+            {
+                var rng = np.random.RandomState(seed);
+                rng.shuffle(start_positions);
+            }
+
+            var sequence_length_tensor = math_ops.cast(sequence_length, dtype: index_dtype);
+            var sampling_rate_tensor = math_ops.cast(sampling_rate, dtype: index_dtype);
+
+            var start_positions_tensor = tf.constant(start_positions);
+            var positions_ds = tf.data.Dataset.from_tensor(start_positions_tensor).repeat();
+            var z = tf.data.Dataset.zip(tf.data.Dataset.range(len(start_positions)), positions_ds);
+            var indices = z.map(m =>
+            {
+                var (i, positions) = (m[0], m[1]);
+                return tf.range(positions[i], positions[i] + sequence_length_tensor * sampling_rate_tensor, sampling_rate_tensor);
+            }, num_parallel_calls: -1);
+            return null;
+        }
+
+        IDatasetV2 sequences_from_indices(Tensor array, Tensor indices_ds, Tensor start_index, Tensor end_index)
+        {
+            return null;
+        }
     }
 }

src/TensorFlowNET.Keras/Tensorflow.Keras.csproj

@@ -6,7 +6,7 @@
     <LangVersion>8.0</LangVersion>
     <RootNamespace>Tensorflow.Keras</RootNamespace>
     <Platforms>AnyCPU;x64</Platforms>
-    <Version>0.5.1</Version>
+    <Version>0.5.2</Version>
     <Authors>Haiping Chen</Authors>
     <Product>Keras for .NET</Product>
     <Copyright>Apache 2.0, Haiping Chen 2021</Copyright>
@@ -24,7 +24,8 @@
 * Support model.load_weights.
 * Add Subtract layer
 * Support YOLOv3 model.
-* Text preprocessing</PackageReleaseNotes>
+* Text preprocessing
+* Preprocessing.timeseries_dataset_from_array</PackageReleaseNotes>
     <Description>Keras for .NET
 
 Keras is an API designed for human beings, not machines. Keras follows best practices for reducing cognitive load: it offers consistent &amp; simple APIs, it minimizes the number of user actions required for common use cases, and it provides clear &amp; actionable error messages.</Description>
@@ -35,8 +36,8 @@ Keras is an API designed for human beings, not machines. Keras follows best prac
     <RepositoryType>Git</RepositoryType>
     <SignAssembly>true</SignAssembly>
     <AssemblyOriginatorKeyFile>Open.snk</AssemblyOriginatorKeyFile>
-    <AssemblyVersion>0.5.1.0</AssemblyVersion>
-    <FileVersion>0.5.1.0</FileVersion>
+    <AssemblyVersion>0.5.2.0</AssemblyVersion>
+    <FileVersion>0.5.2.0</FileVersion>
     <PackageLicenseFile>LICENSE</PackageLicenseFile>
   </PropertyGroup>