Fix error in semrel training, mapping feats as features and finetuning paths handling.

combo/data/token_indexers/token_features_indexer.py

@@ -52,7 +52,11 @@ class TokenFeatsIndexer(data.TokenIndexer):
             if feat in ["_", "__ROOT__"]:
                 pass
             else:
+                # Handle case where feature is binary (doesn't have associated value)
+                if value:
                     features.append(feat + "=" + value)
+                else:
+                    features.append(feat)
         return features
 
     @overrides

combo/main.py

@@ -56,10 +56,10 @@ flags.DEFINE_boolean(name="tensorboard", default=False,
                      help="When provided model will log tensorboard metrics.")
 
 # Finetune after training flags
-flags.DEFINE_string(name="finetuning_training_data_path", default="",
-                    help="Training data path")
-flags.DEFINE_string(name="finetuning_validation_data_path", default="",
-                    help="Validation data path")
+flags.DEFINE_list(name="finetuning_training_data_path", default="",
+                  help="Training data path(s)")
+flags.DEFINE_list(name="finetuning_validation_data_path", default="",
+                  help="Validation data path(s)")
 flags.DEFINE_string(name="config_path", default="config.template.jsonnet",
                     help="Config file path.")
 
@@ -101,7 +101,8 @@ def run(_):
         logger.info(f"Training model stored in: {serialization_dir}")
 
         if FLAGS.finetuning_training_data_path:
-            checks.file_exists(FLAGS.finetuning_training_data_path)
+            for f in FLAGS.finetuning_training_data_path:
+                checks.file_exists(f)
 
             # Loading will be performed from stored model.tar.gz
             del model
@@ -171,9 +172,9 @@ def _get_ext_vars(finetuning: bool = False) -> Dict:
         return {}
     return {
         "training_data_path": (
-            ":".join(FLAGS.training_data_path) if not finetuning else FLAGS.finetuning_training_data_path),
+            ",".join(FLAGS.training_data_path if not finetuning else FLAGS.finetuning_training_data_path)),
         "validation_data_path": (
-            ":".join(FLAGS.validation_data_path) if not finetuning else FLAGS.finetuning_validation_data_path),
+            ",".join(FLAGS.validation_data_path if not finetuning else FLAGS.finetuning_validation_data_path)),
         "pretrained_tokens": FLAGS.pretrained_tokens,
         "pretrained_transformer_name": FLAGS.pretrained_transformer_name,
         "features": " ".join(FLAGS.features),

combo/models/parser.py

@@ -42,7 +42,11 @@ class HeadPredictionModel(base.Predictor):
             lengths = mask.data.sum(dim=1).long().cpu().numpy() + 1
             for idx, length in enumerate(lengths):
                 probs = x[idx, :].softmax(dim=-1).cpu().numpy()
-                probs[:, 0] = 0
+
+                # We do not want any word to be parent of the root node (ROOT, 0).
+                # Also setting it to -1 instead of 0 fixes edge case where softmax made all
+                # but ROOT prediction to EXACTLY 0.0 and it might cause in many ROOT -> word edges)
+                probs[:, 0] = -1
                 heads, _ = chu_liu_edmonds.decode_mst(probs.T, length=length, has_labels=False)
                 heads[0] = 0
                 pred.append(heads)

config.template.jsonnet

@@ -58,6 +58,7 @@ local loss_weights = {
     feats: 0.2,
     deprel: 0.8,
     head: 0.2,
+    semrel: 0.05,
 };
 # Encoder hidden size, int
 local hidden_size = 512;

tests/fixtures/example.conllu

 # sent_id = test-s1
 # text = Easy sentence.
-1	Verylongwordwhichmustbetruncatedbythesystemto30	easy	ADJ	adj	AdpType=Prep	1	amod	_	_
+1	Verylongwordwhichmustbetruncatedbythesystemto30	easy	ADJ	adj	AdpType=Prep|Adp	1	amod	_	_
 2	Sentence	verylonglemmawhichmustbetruncatedbythesystemto30	NOUN	nom	Number=Sing	0	root	_	_
 3	.	.	PUNCT	.	_	1	punct	_	_