revert

python/mxnet/gluon/rnn/rnn_layer.py

@@ -180,6 +180,10 @@ def begin_state(self, batch_size=0, func=ndarray.zeros, **kwargs):
     def hybrid_forward(self, F, inputs, states=None, **kwargs):
         if F is ndarray:
             batch_size = inputs.shape[self._layout.find('N')]
+            if self._input_size == 0:
+                for i in range(self._dir):
+                    self.i2h_weight[i].shape = (self._gates*self._hidden_size, inputs.shape[2])
+                    self.i2h_weight[i]._finish_deferred_init()
         skip_states = states is None
         if skip_states:
             if F is ndarray:
@@ -199,12 +203,13 @@ def hybrid_forward(self, F, inputs, states=None, **kwargs):
         # out is (output, state)
         return out[0] if skip_states else out
 
-    def infer_shape(self, inputs, *states):
-        if self._input_size == 0:
+    def __call__(self, inputs, *states):
+        if self._input_size == 0 and isinstance(inputs, ndarray.NDArray):
             for i in range(self._dir):
                 self.i2h_weight[i].shape = (self._gates*self._hidden_size, inputs.shape[2])
                 self.i2h_weight[i]._finish_deferred_init()
-        return super(_RNNLayer, self).infer_shape(inputs, *states)
+        states = list(filter(lambda x: x is not None, states))
+        return super(_RNNLayer, self).__call__(inputs, *states)
 
     def _forward_kernel(self, F, inputs, states, **kwargs):
         """ forward using CUDNN or CPU kenrel"""