cntk backend: fix the reversed rnn bug (keras-team#7593)

* fix the reversed rnn bug * udpate error message. * Fix error msg
shaform · Aug 11, 2017 · 3537381 · 3537381
1 parent c2b844b
commit 3537381
Showing 1 changed file with 11 additions and 4 deletions.
diff --git a/keras/backend/cntk_backend.py b/keras/backend/cntk_backend.py
@@ -1310,7 +1310,15 @@ def rnn(step_function, inputs, initial_states,
             initial.append(s)
 
     need_convert = not has_seq_axis(inputs)
+    if go_backwards and need_convert is False:
+        raise NotImplementedError('CNTK Backend: `go_backwards` is not supported with '
+                                  'variable-length sequences. Please specify a '
+                                  'static length for your sequences.')
+
     if need_convert:
+        if go_backwards:
+            inputs = reverse(inputs, 1)
+
         inputs = C.to_sequence(inputs)
 
         j = 0
@@ -1327,6 +1335,8 @@ def rnn(step_function, inputs, initial_states,
             j += 1
 
     if mask is not None and not has_seq_axis(mask):
+        if go_backwards:
+            mask = reverse(mask, 1)
         if len(int_shape(mask)) == 2:
             mask = expand_dims(mask)
         mask = C.to_sequence_like(mask, inputs)
@@ -1339,10 +1349,7 @@ def _recurrence(x, states, m):
             place_holders = [C.placeholder(dynamic_axes=x.dynamic_axes) for _ in states]
             past_values = []
             for s, p in zip(states, place_holders):
-                past_values.append(
-                    C.sequence.past_value(
-                        p, s) if go_backwards is False else C.sequence.future_value(
-                        p, s))
+                past_values.append(C.sequence.past_value(p, s))
             new_output, new_states = step_function(
                 x, tuple(past_values) + tuple(constants))
             if m is not None: