Fix some recursive functions (e.g., reorder_incremental_state) to onl…

…y touch each module once (facebookresearch#379)

Summary:
This can happen if a module is registered in more than one place in the network.
Pull Request resolved: facebookresearch#379

Differential Revision: D13154498

Pulled By: myleott

fbshipit-source-id: a35575d1956a46cd35ac8b16a719ad20ac3e380a

fairseq/models/fairseq_incremental_decoder.py

@@ -56,19 +56,26 @@ def reorder_incremental_state(self, incremental_state, new_order):
         previous time step. A typical use case is beam search, where the input
         order changes between time steps based on the selection of beams.
         """
+        seen = set()
+
         def apply_reorder_incremental_state(module):
-            if module != self and hasattr(module, 'reorder_incremental_state'):
-                module.reorder_incremental_state(
-                    incremental_state,
-                    new_order,
-                )
+            if module != self and hasattr(module, 'reorder_incremental_state') \
+                    and module not in seen:
+                seen.add(module)
+                module.reorder_incremental_state(incremental_state, new_order)
+
         self.apply(apply_reorder_incremental_state)
 
     def set_beam_size(self, beam_size):
         """Sets the beam size in the decoder and all children."""
         if getattr(self, '_beam_size', -1) != beam_size:
+            seen = set()
+
             def apply_set_beam_size(module):
-                if module != self and hasattr(module, 'set_beam_size'):
+                if module != self and hasattr(module, 'set_beam_size') \
+                        and module not in seen:
+                    seen.add(module)
                     module.set_beam_size(beam_size)
+
             self.apply(apply_set_beam_size)
             self._beam_size = beam_size

fairseq/models/fairseq_model.py

@@ -99,8 +99,12 @@ def apply_remove_weight_norm(module):
 
         self.apply(apply_remove_weight_norm)
 
+        seen = set()
+
         def apply_make_generation_fast_(module):
-            if module != self and hasattr(module, 'make_generation_fast_'):
+            if module != self and hasattr(module, 'make_generation_fast_') \
+                    and module not in seen:
+                seen.add(module)
                 module.make_generation_fast_(**kwargs)
 
         self.apply(apply_make_generation_fast_)
@@ -115,8 +119,12 @@ def train(mode):
 
     def prepare_for_onnx_export_(self, **kwargs):
         """Make model exportable via ONNX trace."""
+        seen = set()
+
         def apply_prepare_for_onnx_export_(module):
-            if module != self and hasattr(module, 'prepare_for_onnx_export_'):
+            if module != self and hasattr(module, 'prepare_for_onnx_export_') \
+                    and module not in seen:
+                seen.add(module)
                 module.prepare_for_onnx_export_(**kwargs)
 
         self.apply(apply_prepare_for_onnx_export_)