RNNG: add tok/beam size to forward params (facebookresearch#434)

Summary:
Pull Request resolved: facebookresearch#434

1 add tok/beam size to forward params
2 change the way of getting input from using enum to incrementing index, because the enum way is wrong when dict feat is optional and contextual feat is given (input position changed)
3 this change is backward compatible, but not forward compatible (old model will work with this code but new model will not work with existing predictor)

Differential Revision: D14718588

fbshipit-source-id: 2289c97840a3e77d0a8f807f2d558be9b98da710

pytext/data/compositional_data_handler.py

@@ -146,6 +146,8 @@ def _train_input_from_batch(self, batch):
             if name == ACTION_FEATURE_FIELD:
                 input = input.tolist()  # Action needn't be passed as Tensor obj.
             m_inputs.append(input)
+        # beam size and topk
+        m_inputs.extend([1, 1])
         return m_inputs
 
     def _test_input_from_batch(self, batch):
@@ -156,6 +158,8 @@ def _test_input_from_batch(self, batch):
             getattr(batch, DatasetFieldName.DICT_FIELD, None),
             None,
             getattr(batch, DatasetFieldName.PRETRAINED_MODEL_EMBEDDING, None),
+            1,
+            1,
         ]
 
     def preprocess_row(self, row_data: Dict[str, Any]) -> Dict[str, Any]:

pytext/models/semantic_parsers/rnng/rnng_parser.py

@@ -91,6 +91,7 @@ class RNNGConstraints(ConfigBase):
         # version 0 - initial implementation
         # version 1 - beam search
         # version 2 - use zero init state rather than random
+        # version 3 - add beam search input params
         version: int = 2
         lstm: BiLSTM.Config = BiLSTM.Config()
         ablation: AblationParams = AblationParams()
@@ -125,8 +126,6 @@ def from_config(cls, model_config, feature_config, metadata: CommonMetadata):
             lstm_dim=model_config.lstm.lstm_dim,
             max_open_NT=model_config.max_open_NT,
             dropout=model_config.dropout,
-            beam_size=model_config.beam_size,
-            top_k=model_config.top_k,
             actions_vocab=metadata.actions_vocab,
             shift_idx=metadata.shift_idx,
             reduce_idx=metadata.reduce_idx,
@@ -146,8 +145,6 @@ def __init__(
         lstm_dim: int,
         max_open_NT: int,
         dropout: float,
-        beam_size: int,
-        top_k: int,
         actions_vocab,
         shift_idx: int,
         reduce_idx: int,
@@ -238,9 +235,6 @@ def __init__(
         self.valid_IN_idxs = valid_IN_idxs
         self.valid_SL_idxs = valid_SL_idxs
 
-        self.beam_size = beam_size
-        self.top_k = top_k
-
         num_actions = len(actions_vocab)
         lstm_count = ablation.use_buffer + ablation.use_stack + ablation.use_action
         if lstm_count == 0:
@@ -278,6 +272,8 @@ def forward(
         dict_feat: Optional[Tuple[torch.Tensor, ...]] = None,
         actions: Optional[List[List[int]]] = None,
         contextual_token_embeddings: Optional[torch.Tensor] = None,
+        beam_size=1,
+        top_k=1,
     ) -> List[Tuple[torch.Tensor, torch.Tensor]]:
         """RNNG forward function.
 
@@ -295,8 +291,6 @@ def forward(
             (batch_size, action_length)
             (batch_size, action_length, number_of_actions)
         """
-        beam_size = self.beam_size
-        top_k = self.top_k
 
         if self.stage != Stage.TEST:
             beam_size = 1

pytext/models/test/rnng_test.py

@@ -79,8 +79,6 @@ def setUp(self):
             lstm_dim=20,
             max_open_NT=10,
             dropout=0.2,
-            beam_size=3,
-            top_k=3,
             actions_vocab=actions_vocab,
             shift_idx=4,
             reduce_idx=3,
@@ -218,7 +216,9 @@ def test_forward_shapes(self):
 
         # Beam Search Test
         self.parser.eval(Stage.TEST)
-        results = self.parser(tokens=tokens, seq_lens=seq_lens, dict_feat=dict_feat)
+        results = self.parser(
+            tokens=tokens, seq_lens=seq_lens, dict_feat=dict_feat, beam_size=3, top_k=3
+        )
         self.assertEqual(len(results), 3)
         for actions, scores in results:
             self.assertGreater(actions.shape[1], tokens.shape[1])