Decouple decoder and output layer creation in BasePairwiseModel (#973)

Summary:
Pull Request resolved: #973

We create both the decoder and output layer in `BasePairwiseModel._create_decoder()`. This diff moves output layer creation out of `_create_decoder()` function.

Reviewed By: borguz

Differential Revision: D17318025

fbshipit-source-id: 556c988831d6dcf0788c1a61efebfb8f8e815940

pytext/models/bert_classification_models.py

@@ -170,8 +170,9 @@ def _create_encoder(
     @classmethod
     def from_config(cls, config: Config, tensorizers: Dict[str, Tensorizer]):
         encoder1, encoder2 = cls._create_encoder(config, tensorizers)
-        decoder, output_layer = cls._create_decoder(
-            config, [encoder1, encoder2], tensorizers
+        decoder = cls._create_decoder(config, [encoder1, encoder2], tensorizers)
+        output_layer = create_module(
+            config.output_layer, labels=tensorizers["labels"].vocab
         )
         return cls(encoder1, encoder2, decoder, output_layer, config.encode_relations)
 

pytext/models/pair_classification_model.py

@@ -76,8 +76,7 @@ def _create_decoder(
         decoder = create_module(
             config.decoder, in_dim=decoder_in_dim, out_dim=len(labels)
         )
-        output_layer = create_module(config.output_layer, labels=labels)
-        return decoder, output_layer
+        return decoder
 
     @classmethod
     def _encode_relations(cls, encodings: List[torch.Tensor]) -> List[torch.Tensor]:
@@ -205,8 +204,9 @@ def _create_representations(cls, config: Config, embeddings: nn.ModuleList):
     def from_config(cls, config: Config, tensorizers: Dict[str, Tensorizer]):
         embeddings = cls._create_embeddings(config, tensorizers)
         representations = cls._create_representations(config, embeddings)
-        decoder, output_layer = cls._create_decoder(
-            config, representations, tensorizers
+        decoder = cls._create_decoder(config, representations, tensorizers)
+        output_layer = create_module(
+            config.output_layer, labels=tensorizers["labels"].vocab
         )
         return cls(
             embeddings, representations, decoder, output_layer, config.encode_relations