fix: temporarily fix the sequential nest error.

src/TensorFlowNET.Core/Training/Saving/SavedModel/save.cs

@@ -88,7 +88,7 @@ public static partial class SavedModelUtils
         {
             if (ops.inside_function())
             {
-                throw new AssertionError("`tf.saved_model.save` is not supported inside a traced @tf.function. " +
+                throw new AssertionError("`tf.saved_model.save` is not supported inside a traced [AutoGraph]. " +
                                          "Move the call to the outer eagerly-executed context.");
             }
 

src/TensorFlowNET.Keras/Engine/Layer.Apply.cs

@@ -41,5 +41,19 @@ namespace Tensorflow.Keras.Engine
 
             return outputs;
         }
+
+        // TODO(Rinne): remove it and completely fix issue 1084
+        [Obsolete]
+        private bool _enforce_layer_construction = false;
+        [Obsolete]
+        internal void enforce_layer_construction()
+        {
+            _enforce_layer_construction = true;
+        }
+        [Obsolete]
+        internal void unset_layer_construction()
+        {
+            _enforce_layer_construction = false;
+        }
     }
 }

src/TensorFlowNET.Keras/Engine/Layer.cs

@@ -291,7 +291,7 @@ namespace Tensorflow.Keras.Engine
         bool _in_functional_construction_mode(Tensors inputs)
         {
             return tf.Context.executing_eagerly()
-                && inputs.Count(x => x is not EagerTensor && x is not NDArray) == inputs.Count();
+                && inputs.Count(x => x is not EagerTensor && x is not NDArray) == inputs.Count() || _enforce_layer_construction;
         }
 
         public void SetConnectivityMetadata(Tensors inputs, Tensors outputs)

src/TensorFlowNET.Keras/Engine/Sequential.cs

@@ -62,7 +62,17 @@ namespace Tensorflow.Keras.Engine
         {
             foreach(var layer in layers)
             {
+                // TODO(Rinne): remove it and completely fix issue 1084
+                if(layer is Sequential s)
+                {
+                    s.Layers.ForEach(x => ((Layer)x).enforce_layer_construction());
+                }
                 add(layer);
+                // TODO(Rinne): remove it and completely fix issue 1084
+                if (layer is Sequential s2)
+                {
+                    s2.Layers.ForEach(x => ((Layer)x).unset_layer_construction());
+                }
             }
         }
 
@@ -163,7 +173,7 @@ namespace Tensorflow.Keras.Engine
             Tensors layer_output = null;
             Tensors outputs = null;
             List<INode> created_nodes = new List<INode>();
-            foreach (var layer in args.Layers)
+            foreach (var layer in Layers)
             {
                 clear_previously_created_nodes(layer, _created_nodes);
                 layer_output = layer.Apply(layer_input);

test/TensorFlowNET.Keras.UnitTest/Model/ModelBuildTest.cs

@@ -1,5 +1,7 @@
 using Microsoft.VisualStudio.TestTools.UnitTesting;
+using System;
 using static Tensorflow.Binding;
+using static Tensorflow.KerasApi;
 
 namespace Tensorflow.Keras.UnitTest.Model
 {
@@ -14,24 +16,47 @@ namespace Tensorflow.Keras.UnitTest.Model
             var dense = tf.keras.layers.Dense(64);
             var output = dense.Apply(input);
             var model = tf.keras.Model(input, output);
+            model.compile(tf.keras.optimizers.Adam(), tf.keras.losses.CategoricalCrossentropy());
 
             // one dimensions input with unknown batchsize
             var input_2 = tf.keras.layers.Input((60));
             var dense_2 = tf.keras.layers.Dense(64);
-            var output_2 = dense.Apply(input_2);
+            var output_2 = dense_2.Apply(input_2);
             var model_2 = tf.keras.Model(input_2, output_2);
+            model_2.compile(tf.keras.optimizers.Adam(), tf.keras.losses.CategoricalCrossentropy());
 
             // two dimensions input with specified batchsize
             var input_3 = tf.keras.layers.Input((17, 60), 8);
             var dense_3 = tf.keras.layers.Dense(64);
-            var output_3 = dense.Apply(input_3);
+            var output_3 = dense_3.Apply(input_3);
             var model_3 = tf.keras.Model(input_3, output_3);
+            model_3.compile(tf.keras.optimizers.Adam(), tf.keras.losses.CategoricalCrossentropy());
 
             // one dimensions input with specified batchsize
             var input_4 = tf.keras.layers.Input((60), 8);
             var dense_4 = tf.keras.layers.Dense(64);
-            var output_4 = dense.Apply(input_4);
+            var output_4 = dense_4.Apply(input_4);
             var model_4 = tf.keras.Model(input_4, output_4);
+            model_4.compile(tf.keras.optimizers.Adam(), tf.keras.losses.CategoricalCrossentropy());
+        }
+
+        [TestMethod]
+        public void NestedSequential()
+        {
+            var block1 = keras.Sequential(new[] { 
+                keras.layers.InputLayer((3, 3)), 
+                keras.Sequential(new []
+                    {
+                        keras.layers.Flatten(),
+                        keras.layers.Dense(5)
+                    }
+                )
+            });
+            block1.compile(tf.keras.optimizers.Adam(), tf.keras.losses.CategoricalCrossentropy());
+
+            var x = tf.ones((1, 3, 3));
+            var y = block1.predict(x);
+            Console.WriteLine(y);
         }
     }
 }