Fixed a cloning bug in placehoder shape information

Source/CNTKv2LibraryDll/Function.cpp

@@ -805,7 +805,9 @@ namespace CNTK
 
                             if (existingPlaceholderReplacement == placeholderReplacements.end())
                             {
-                                clonedInput = PlaceholderVariable();
+                                //we need to carry the shape information to the new placeholder otherwise, deep chained recurrence with reshaping ops will fail (e.g. expand_dims);
+                                //however, we can not carry over the dynamic axis, as the placeholder might be replaced with different dynamic axes
+                                clonedInput = PlaceholderVariable(cloneeInput.Shape());
                                 placeholderReplacements[clonedInput] = cloneeInput;
                             }
                             else

bindings/python/cntk/ops/tests/function_tests.py

@@ -558,3 +558,23 @@ def test_clone_with_different_dynamic_axes():
 
     rnn = C.layers.Recurrence(C.layers.LSTM(5))(question_input)
     rnn_cloned = rnn.clone(C.CloneMethod.share, {question_input:answer_input})
+
+
+def test_clone_with_deep_rnn_chaining():
+    def seq_op_func(seqinp):
+        l = seqinp
+        r = C.sequence.future_value(l)
+        r = C.expand_dims(r, -len(seqinp.shape) - 1)
+        res = l + r
+        return res
+
+    def rnn_seq(features):
+        step_func = C.layers.GRU(1)
+        seq = C.layers.Recurrence(step_func)(features)
+        return seq
+
+    feat = C.sequence.input_variable((40,), name='sequence_inp')
+    c1 = rnn_seq(feat)
+    seq_op_res = seq_op_func(c1)
+    net = rnn_seq(seq_op_res)
+    cloned = net.clone('freeze')