v0.12 regression: Fix registration of children for Block

python/mxnet/gluon/block.py

@@ -191,9 +191,10 @@ def __setattr__(self, name, value):
                 for i, c in enumerate(self._children):
                     if c is existing:
                         self._children[i] = value
-        else:
-            if isinstance(value, Block):
+            elif isinstance(value, Block):
                 self.register_child(value)
+        elif isinstance(value, Block):
+            self.register_child(value)
 
         super(Block, self).__setattr__(name, value)
 
@@ -332,6 +333,8 @@ def __init__(self, prefix=None, params=None):
     def __setattr__(self, name, value):
         """Registers parameters."""
         super(HybridBlock, self).__setattr__(name, value)
+        if isinstance(value, HybridBlock):
+            self._clear_cached_op()
         if isinstance(value, Parameter):
             assert name not in self._reg_params or \
                 not isinstance(self._reg_params[name], Parameter), \

tests/python/unittest/test_gluon.py

@@ -516,6 +516,20 @@ def test_hybrid_stale_cache():
     net.add(mx.gluon.nn.Flatten())
     assert net(mx.nd.ones((2,3,5))).shape == (2, 30)
 
+    net = mx.gluon.nn.HybridSequential()
+    with net.name_scope():
+        net.fc1 = mx.gluon.nn.Dense(10, weight_initializer='zeros',
+                                    bias_initializer='ones', flatten=False)
+        net.fc2 = mx.gluon.nn.Dense(10, weight_initializer='zeros',
+                                    bias_initializer='ones', flatten=False)
+    net.hybridize()
+    net.initialize()
+    net(mx.nd.ones((2,3,5)))
+
+    net.fc2 = mx.gluon.nn.Dense(10, weight_initializer='zeros',
+                                bias_initializer='ones', flatten=True)
+    net.initialize()
+    assert net(mx.nd.ones((2,3,5))).shape == (2, 10)
 
 if __name__ == '__main__':
     import nose