fix AdamOptimizer for Graph mode.

Author: Oceania2018

src/TensorFlowNET.Core/Gradients/math_grad.cs

@@ -542,15 +542,28 @@ public static Tensor[] _SumGrad(Operation op, Tensor[] grads)
             }
 
             input_shape = array_ops.shape(op.inputs[0]);
-            if (!op.get_attr<bool>("keep_dims"))
+
+            if (tf.executing_eagerly())
+            {
+                if (!op.get_attr<bool>("keep_dims"))
+                {
+                    ops.colocate_with(input_shape);
+                    var output_shape_kept_dims = math_ops.reduced_shape(input_shape, op.inputs[1]);
+                    // var tile_scaling = _safe_shape_div(input_shape, output_shape_kept_dims);
+                    grad = gen_array_ops.reshape(grad, output_shape_kept_dims);
+                }
+
+                return new Tensor[] { gen_array_ops.broadcast_to(grad, input_shape), null };
+            }
+            else
             {
                 ops.colocate_with(input_shape);
                 var output_shape_kept_dims = math_ops.reduced_shape(input_shape, op.inputs[1]);
-                // var tile_scaling = _safe_shape_div(input_shape, output_shape_kept_dims);
+                var tile_scaling = _safe_shape_div(input_shape, output_shape_kept_dims);
                 grad = gen_array_ops.reshape(grad, output_shape_kept_dims);
-            }
 
-            return new Tensor[] { gen_array_ops.broadcast_to(grad, input_shape), null };
+                return new Tensor[] { gen_array_ops.tile(grad, tile_scaling), null };
+            }
         }
 
         [RegisterGradient("RealDiv")]

src/TensorFlowNET.Core/Operations/gen_image_ops.cs

@@ -66,7 +66,7 @@ public static Tensor decode_jpeg(Tensor contents,
             int ratio = 1,
             bool fancy_upscaling = true,
             bool try_recover_truncated = false,
-            int acceptable_fraction = 1,
+            float acceptable_fraction = 1,
             string dct_method = "",
             string name = null)
         {

src/TensorFlowNET.Core/Operations/math_ops.cs

@@ -652,7 +652,7 @@ private static Tensor _ReductionDims(Tensor x, Tensor axis)
             }
             else
             {
-                if(x.rank > -1)
+                if (x.rank > -1 && tf.executing_eagerly())
                     return constant_op.constant(np.arange(x.rank));
 
                 var rank = array_ops.rank(x);

src/TensorFlowNET.Core/Training/AdamOptimizer.cs

@@ -109,7 +109,7 @@ private Operation _apply_sparse_shared(Tensor grad, IVariableV1 var, Tensor indi
             return control_flow_ops.group(new[] { var_update, m_t, v_t });
         }
 
-        protected override void _create_slots(ResourceVariable[] var_list)
+        protected override void _create_slots(IVariableV1[] var_list)
         {
             var first_var = var_list.OrderBy(x => x.Name).First();
             _create_non_slot_variable(initial_value: _beta1, name: "beta1_power", colocate_with: first_var);

src/TensorFlowNET.Core/Training/Optimizer.cs

@@ -107,7 +107,7 @@ public Optimizer(Tensor learning_rate, bool use_locking, string name = null)
         /// </returns>
         public Operation minimize(Tensor loss, 
             IVariableV1 global_step = null,
-            List<ResourceVariable> var_list=null,
+            List<IVariableV1> var_list=null,
             GateGradientType gate_gradients = GateGradientType.GATE_OP,
             int? aggregation_method=null,
             bool colocate_gradients_with_ops = false, string name=null, Tensor grad_loss=null)
@@ -142,17 +142,17 @@ public Operation minimize(Tensor loss,
         /// <returns>
         /// An `Operation` that applies the specified gradients. If `global_step`
         /// was not None, that operation also increments `global_step`.</returns>
-        public Operation apply_gradients(Tuple<Tensor, ResourceVariable>[] grads_and_vars, IVariableV1 global_step = null, string name = null)
+        public Operation apply_gradients(Tuple<Tensor, IVariableV1>[] grads_and_vars, IVariableV1 global_step = null, string name = null)
         {
             // No DistributionStrategy case.
-            var converted_grads_and_vars = new List<(Tensor, ResourceVariable, _OptimizableVariable)>();
+            var converted_grads_and_vars = new List<(Tensor, IVariableV1, _OptimizableVariable)>();
             foreach (var (g, v) in grads_and_vars)
             {
                 if(g != null)
                 {
                     // Convert the grad to Tensor or IndexedSlices if necessary.
                     var gR = ops.convert_to_tensor_or_indexed_slices(g);
-                    var p = optimizer._get_processor(v);
+                    var p = optimizer._get_processor(v as ResourceVariable);
                     converted_grads_and_vars.Add((gR, v, p));
                 }
             }
@@ -230,7 +230,7 @@ public Operation apply_gradients(Tuple<Tensor, ResourceVariable>[] grads_and_var
         /// silently ignored).
         /// </summary>
         /// <param name="var_list"></param>
-        protected virtual void _create_slots(ResourceVariable[] var_list)
+        protected virtual void _create_slots(IVariableV1[] var_list)
         {
 
         }
@@ -369,8 +369,8 @@ protected IVariableV1 _get_non_slot_variable(string name, Graph graph = null)
         /// A list of (gradient, variable) pairs. Variable is always present, but
         /// gradient can be `None`.
         /// </returns>
-        public Tuple<Tensor, ResourceVariable>[] compute_gradients(Tensor loss,
-            List<ResourceVariable> var_list = null,
+        public Tuple<Tensor, IVariableV1>[] compute_gradients(Tensor loss,
+            List<IVariableV1> var_list = null,
             int? aggregation_method = null,
             GateGradientType gate_gradients = GateGradientType.GATE_OP,
             bool colocate_gradients_with_ops = false,
@@ -381,26 +381,13 @@ public Tuple<Tensor, ResourceVariable>[] compute_gradients(Tensor loss,
 
             if(var_list == null)
             {
-                var vars = ops.get_collection<ResourceVariable>(tf.GraphKeys.TRAINABLE_RESOURCE_VARIABLES);
+                var vars = ops.get_collection<IVariableV1>(tf.GraphKeys.TRAINABLE_RESOURCE_VARIABLES);
                 var tmp = variables.trainable_variables();
-                switch (tmp)
-                {
-                    case List<ResourceVariable> values:
-                        var_list = values.Concat(vars).ToList();
-                        break;
-                    /*case List<RefVariable> values:
-                        var_list = values.Concat(vars).ToList();
-                        break;
-                    case List<IVariableV1> values:
-                        var_list = values.Select(x => x as RefVariable).Concat(vars).ToList();
-                        break;*/
-                    default:
-                        throw new NotImplementedException("");
-                }
+                var_list = (tmp as List<IVariableV1>).Concat(vars).ToList();
             }
 
-            var_list = var_list.Concat(ops.get_collection<ResourceVariable>(tf.GraphKeys._STREAMING_MODEL_PORTS)).ToList();
-            var processors = var_list.Select(v => optimizer._get_processor(v)).ToList();
+            var_list = var_list.Concat(ops.get_collection<IVariableV1>(tf.GraphKeys._STREAMING_MODEL_PORTS)).ToList();
+            var processors = var_list.Select(v => optimizer._get_processor(v as ResourceVariable)).ToList();
             var var_refs = processors.Select(x => x.target()).ToArray();
 
             var grads = gradients_impl.gradients(new Tensor[] { loss }, var_refs, grad_ys: grad_loss == null ? null : new Tensor[] { grad_loss },
@@ -412,7 +399,7 @@ public Tuple<Tensor, ResourceVariable>[] compute_gradients(Tensor loss,
                 grads = control_flow_ops.tuple(grads);
 
             var grads_and_vars = zip(grads, var_list)
-                .Select(x => new Tuple<Tensor, ResourceVariable>(x.Item1, x.Item2))
+                .Select(x => new Tuple<Tensor, IVariableV1>(x.Item1, x.Item2))
                 .ToArray();
 
             return grads_and_vars;

src/TensorFlowNET.Core/Variables/ResourceVariable.cs

@@ -155,7 +155,7 @@ private void _init_from_args(object initial_value = null,
                         _graph_element = value;
                     });
 
-                    ops.add_to_collections(collections, this);
+                    ops.add_to_collections<IVariableV1>(collections, this);
                 }
                 else
                 {
@@ -184,6 +184,8 @@ private void _init_from_proto(VariableDef variable_def, string import_scope = nu
             var g = ops.get_default_graph();
             var prepend_name_scope = ops.prepend_name_scope(variable_def.VariableName, import_scope: import_scope);
             handle = g.as_graph_element(prepend_name_scope) as Tensor;
+            _handle_name = handle.name;
+            _name = handle.name;
             _shape = new TensorShape(handle.op.get_attr("shape") as TensorShapeProto);
 
             prepend_name_scope = ops.prepend_name_scope(variable_def.InitializerName, import_scope: import_scope);