Restricting loops to Delay, PastValue, and FutureValue

contrib/Python/cntk/graph.py

@@ -140,6 +140,17 @@ class ComputationNode(object):
 
         return param
 
+
+    def _is_forward_ref(self, p_name, p_value): 
+        '''
+        Although the unrolled graph is a DAG, when we specify recurrence we
+        naturally have loops. We can resolve this by using forward references.
+        This method is checking whether the particular name and value of this
+        instance are actually one of those forward references.
+        '''
+        is_loop_node = self.name in ('Delay', 'PastValue', 'FutureValue')
+        return is_loop_node and p_name == 'input' and isinstance(p_value, str)
+
     def _to_config_recursively(self, desc, unrolled_nodes, inputs,
                                readers, node_counter=0):
         param_variable_names = []
@@ -161,8 +172,8 @@ class ComputationNode(object):
                     input_nodes_vars = []
                     for pv in inputs_param:
                         if pv in unrolled_nodes:
-                            # we have seen this node already, so just retrieve its
-                            # name
+                            # We have seen this node already, so just retrieve its
+                            # name.
                             child_var = unrolled_nodes[pv]
                         else:
                             child_var, node_counter, child_desc = pv._to_config_recursively(
@@ -173,7 +184,7 @@ class ComputationNode(object):
                     param_variable_names.append(
                         _tuple_to_cntk_shape(input_nodes_vars))
                 else:
-                    if p_name == 'input' and isinstance(p_value, str):
+                    if self._is_forward_ref(p_name, p_value):
                         # We have a forward reference to a node that will be
                         # later on defined. p_value is the var_name of the
                         # later defined node.