fixed Kernel weight wasn't updated for Keras Dense layer #639

src/TensorFlowNET.Core/Binding.cs

@@ -1,5 +1,6 @@
 using System;
 using System.Collections.Generic;
+using System.Diagnostics;
 using System.Dynamic;
 using System.Text;
 
@@ -7,6 +8,7 @@ namespace Tensorflow
 {
     public static partial class Binding
     {
+        [DebuggerNonUserCode]
         public static tensorflow tf { get; } = New<tensorflow>();
 
         /// <summary>

src/TensorFlowNET.Core/Contexts/Context.cs

@@ -65,6 +65,7 @@ namespace Tensorflow.Contexts
         /// Checks whether the current thread has eager execution enabled.
         /// </summary>
         /// <returns></returns>
+        [DebuggerStepThrough]
         public bool executing_eagerly()
             => context_switches.Current().EagerMode;
 

src/TensorFlowNET.Core/Keras/Engine/LossesContainer.cs

@@ -16,7 +16,6 @@ namespace Tensorflow.Keras.Engine
         Mean _loss_metric;
         bool _built;
         Tensor[] _per_output_metrics;
-        List<Tensor> loss_values;
         List<Tensor> loss_metric_values;
 
         public LossesContainer(ILossFunc losses, string[] output_names = null) 
@@ -25,7 +24,6 @@ namespace Tensorflow.Keras.Engine
             _user_losses = losses;
             _losses = losses;
             _loss_metric = new Mean(name: "loss");
-            loss_values = new List<Tensor>();
             loss_metric_values = new List<Tensor>();
             _built = false;
         }
@@ -37,15 +35,17 @@ namespace Tensorflow.Keras.Engine
         /// <param name="y_pred"></param>
         public Tensor Call(Tensor y_true, Tensor y_pred)
         {
-            Build(y_pred);
+            if (!_built)
+                Build(y_pred);
             var loss_value = _losses.Call(y_true, y_pred);
             var loss_metric_value = loss_value;
             var batch_dim = array_ops.shape(y_true)[0];
 
+            var loss_values = new List<Tensor>();
+
             /*if (_losses.Reduction == ReductionV2.SUM_OVER_BATCH_SIZE
                 || _losses.Reduction == ReductionV2.AUTO)
                 loss_value = losses_utils.scale_loss_for_distribution(loss_value);*/
-
             loss_values.append(loss_value);
             loss_metric_values.append(loss_metric_value);
 

src/TensorFlowNET.Core/Keras/Optimizers/RMSprop.cs

@@ -11,6 +11,8 @@ namespace Tensorflow.Keras.Optimizers
     public class RMSprop : OptimizerV2
     {
         RMSpropArgs args;
+        bool centered => args.Centered;
+        protected override string _name => "RMSprop";
 
         public RMSprop(RMSpropArgs args) : base(args)
         {
@@ -18,5 +20,60 @@ namespace Tensorflow.Keras.Optimizers
             _set_hyper("rho", args.RHO);
             _set_hyper("momentum", args.Momentum);
         }
+
+        protected override void _create_slots(IVariableV1[] var_list)
+        {
+            foreach (var var in var_list)
+                add_slot(var, "rms");
+            if (_momentum)
+                foreach (var var in var_list)
+                    add_slot(var, "momentum");
+            if (centered)
+                foreach (var var in var_list)
+                    add_slot(var, "mg");
+        }
+
+        protected override void _prepare_local(DeviceDType device_dtype, Dictionary<DeviceDType, Dictionary<string, Tensor>> _apply_state)
+        {
+            base._prepare_local(device_dtype, _apply_state);
+            var rho = array_ops.identity(_get_hyper("rho", device_dtype.DType));
+            _apply_state[device_dtype]["neg_lr_t"] = -_apply_state[device_dtype]["lr_t"];
+            _apply_state[device_dtype]["epsilon"] = ops.convert_to_tensor(args.Epsilon, dtype: device_dtype.DType);
+            _apply_state[device_dtype]["rho"] = rho;
+            _apply_state[device_dtype]["momentum"] = array_ops.identity(_get_hyper("momentum", device_dtype.DType));
+            _apply_state[device_dtype]["one_minus_rho"] = 1.0f - rho;
+        }
+
+        protected override Operation _resource_apply_dense(IVariableV1 var, Tensor grad, Dictionary<DeviceDType, Dictionary<string, Tensor>> _apply_state)
+        {
+            Dictionary<string, Tensor> coefficients = null;
+            foreach (var state in _apply_state)
+            {
+                if (state.Key.DType == var.dtype.as_base_dtype()
+                    && state.Key.Device == var.Device)
+                {
+                    coefficients = state.Value;
+                    break;
+                }
+            }
+            
+            var rms = get_slot(var, "rms");
+            if (_momentum)
+            {
+                throw new NotImplementedException("");
+            }
+            else
+            {
+                var rms_t = coefficients["rho"] * rms.AsTensor() + coefficients["one_minus_rho"] * math_ops.square(grad);
+                rms_t = state_ops.assign(rms, rms_t, use_locking: _use_locking);
+                var denom_t = rms_t;
+                if (centered)
+                {
+                    throw new NotImplementedException("");
+                }
+                var var_t = var.AsTensor() - coefficients["lr_t"] * grad / (math_ops.sqrt(denom_t) + coefficients["epsilon"]);
+                return state_ops.assign(var, var_t, use_locking: _use_locking).op;
+            }
+        }
     }
 }

src/TensorFlowNET.Core/Tensorflow.Binding.csproj

@@ -25,7 +25,11 @@ https://tensorflownet.readthedocs.io</Description>
 * Eager Mode is added finally.
 * tf.keras is partially working. 
 * tf.data is added.
-* autograph works partially.</PackageReleaseNotes>
+* autograph works partially.
+
+TensorFlow .NET v0.30 is focused on making more Keras API work including:
+* tf.keras.datasets
+* Building keras model in subclass, functional and sequential api</PackageReleaseNotes>
     <FileVersion>0.30.0.0</FileVersion>
     <PackageLicenseFile>LICENSE</PackageLicenseFile>
     <PackageRequireLicenseAcceptance>true</PackageRequireLicenseAcceptance>

src/TensorFlowNET.Core/Tensors/Tensor.cs

@@ -58,7 +58,7 @@ namespace Tensorflow
         ///     The Operation that produces this tensor as an output.
         /// </summary>
         public Operation op => _op;
-        public Tensor[] outputs => op.outputs;
+        public Tensor[] outputs => op?.outputs;
 
         /// <summary>
         /// The string name of this tensor.<br/>