Use operation with customized C API.

src/TensorFlowNET.Core/APIs/c_api.customize.cs

@@ -0,0 +1,13 @@
+using System;
+using System.Collections.Generic;
+using System.Runtime.InteropServices;
+using System.Text;
+
+namespace Tensorflow
+{
+    public partial class c_api
+    {
+        [DllImport(TensorFlowLibName)]
+        public static extern void TFC_SetAttr(SafeGraphHandle graph, IntPtr op, string attr_name, SafeBufferHandle attr_value_proto, SafeStatusHandle status);
+    }
+}

src/TensorFlowNET.Core/Functions/ConcreteFunction.cs

@@ -223,7 +223,7 @@ namespace Tensorflow.Functions
             {
                 input_tangents = new TangentInfo();
             }
-            if(possible_gradient_type == gradients_util.POSSIBLE_GRADIENT_TYPES_FIRST_ORDER)
+            if(possible_gradient_type == gradients_util.POSSIBLE_GRADIENT_TYPES_FIRST_ORDER || tf.Runner.MustRecordGradient())
             {
                 if(input_tangents.Indices is not null || executing_eagerly)
                 {

src/TensorFlowNET.Core/Operations/Operation.cs

@@ -317,27 +317,18 @@ namespace Tensorflow
         {
             Debug.Assert(types.Length == shapes.Length);
             int orig_num_outputs = this.outputs.Length;
-            //var new_outputs = new List<Tensor>(_outputs);
-
-            var old_outputs = _outputs;
-            _outputs = new Tensor[orig_num_outputs + types.Length];
-            for(int i = 0; i <  orig_num_outputs; i++)
-            {
-                _outputs[i] = old_outputs[i];
-            }
+            var new_outputs = new List<Tensor>(_outputs);
 
             // Since the `_outputs` is defined as `Array`, when we add new output, we 
             // have to create a new array, which brings some performance concerns.
             // In the future maybe the type of `outputs` should be reconsidered.
             for(int i = 0; i < types.Length; i++)
             {
-                var t = new Tensor(this, orig_num_outputs + 1, types[i]);
-                _outputs[i] = t;
-                //t = tf.ensure_shape(t, shapes[i]);
+                var t = new Tensor(this, orig_num_outputs + i, types[i]);
                 t.shape = shapes[i];
-                //new_outputs.Add(t);
+                new_outputs.Add(t);
             }
-            //_outputs = new_outputs.ToArray();
+            _outputs = new_outputs.ToArray();
         }
 
         internal void _set_func_attr(string attr_name, string func_name)
@@ -372,23 +363,9 @@ namespace Tensorflow
 
         internal void _set_attr_with_buf(string attr_name, Buffer attr_buf)
         {
-            //if(_op_desc is null)
-            //{
-            //    //var new_node_def = NodeDef.Parser.ParseFrom(node_def.ToByteArray());
-            //    //new_node_def.Name += "_temp";
-            //    //var op = new Operation(new_node_def, graph, inputs, _output_types, control_inputs, _input_types);
-            //    //Status status = new();
-            //    //c_api.TF_SetAttrBool(op._op_desc, "trainable", true);
-            //    ////c_api.TF_SetAttrValueProto(op._op_desc, attr_name, attr_buf.ToArray(), attr_buf.Length, status);
-            //    //status.Check(true);
-            //    // TODO(Rinne): deal with it. Give a warning or make the Operation always contains `op_desc`.
-            //}
-            //else
-            //{
-            //    //Status status = new();
-            //    //c_api.TF_SetAttrValueProto(_op_desc, attr_name, attr_buf.ToArray(), attr_buf.Length, status);
-            //    //status.Check(true);
-            //}
+            Status status = new();
+            c_api.TFC_SetAttr(graph, _handle, attr_name, attr_buf, status);
+            status.Check(true);
         }
     }
 }

src/TensorFlowNET.Core/Tensors/Tensor.cs

@@ -135,7 +135,7 @@ namespace Tensorflow
 
         protected virtual void SetShapeInternal(Shape value)
         {
-            if (value == null)
+            if (value is null || value.ndim == 0 || value.ndim == -1)
                 c_api.TF_GraphSetTensorShape(op.graph.c_graph, _as_tf_output(), null, -1, tf.Status);
             else
                 c_api.TF_GraphSetTensorShape(op.graph.c_graph, _as_tf_output(), value.dims, value.ndim, tf.Status);