More fixes to SentencePiece for T5 (#3515)

* More fixes to spm for T5 Signed-off-by: MaximumEntropy <[email protected]> * Style fixes Signed-off-by: MaximumEntropy <[email protected]>
NVIDIA · Jan 28, 2022 · 33290ac · 33290ac
1 parent 67aae90
commit 33290ac
Show file tree

Hide file tree

Showing 4 changed files with 24 additions and 8 deletions.
diff --git a/nemo/collections/common/tokenizers/sentencepiece_tokenizer.py b/nemo/collections/common/tokenizers/sentencepiece_tokenizer.py
@@ -240,7 +240,10 @@ def unk_id(self):
     @property
     def additional_special_tokens_ids(self):
         """Returns a list of the additional special tokens (excluding bos, eos, pad, unk). Used to return sentinel tokens for e.g. T5."""
-        return list(self.special_token_to_id.values())
+        special_tokens = set(
+            [self.bos_token, self.eos_token, self.pad_token, self.mask_token, self.cls_token, self.sep_token]
+        )
+        return [v for k, v in self.special_token_to_id.items() if k not in special_tokens]
 
     @property
     def vocab(self):

diff --git a/nemo/collections/nlp/data/language_modeling/megatron/t5_dataset.py b/nemo/collections/nlp/data/language_modeling/megatron/t5_dataset.py
@@ -89,7 +89,7 @@ def __init__(
         self.vocab_id_to_token_dict = {idx: token for idx, token in enumerate(self.vocab_id_list)}
 
         self.sentinel_tokens = tokenizer.additional_special_tokens_ids
-        assert len(self.sentinel_tokens) > 0, "Provide the argument --vocab-extra-ids 100 to the script"
+        assert len(self.sentinel_tokens) > 0
 
     def __len__(self):
         return self.samples_mapping.shape[0]

diff --git a/nemo/collections/nlp/models/language_modeling/megatron_t5_model.py b/nemo/collections/nlp/models/language_modeling/megatron_t5_model.py
@@ -480,25 +480,37 @@ def _add_special_tokens_to_tokenizer(self):
 
             # bos, eos, pad and unk may be present in the provided spm .model file, if they are, use it.
             if not hasattr(self.tokenizer, 'pad_token'):
-                if hasattr(self.tokenizer.tokenizer, 'pad_id'):
+                if hasattr(self.tokenizer.tokenizer, 'pad_id') and self.tokenizer.tokenizer.pad_id() > 0:
                     self.tokenizer.pad_token = self.tokenizer.tokenizer.id_to_piece(self.tokenizer.tokenizer.pad_id())
+                else:
+                    self.tokenizer.add_special_tokens({'pad_token': '<pad>'})
             else:
                 self.tokenizer.add_special_tokens({'pad_token': '<pad>'})
 
             if not hasattr(self.tokenizer, 'bos_token'):
-                if hasattr(self.tokenizer.tokenizer, 'bos_id'):
+                if hasattr(self.tokenizer.tokenizer, 'bos_id') and self.tokenizer.tokenizer.bos_id() > 0:
                     self.tokenizer.bos_token = self.tokenizer.tokenizer.id_to_piece(self.tokenizer.tokenizer.bos_id())
+                else:
+                    self.tokenizer.add_special_tokens({'bos_token': '<bos>'})
             else:
                 self.tokenizer.add_special_tokens({'bos_token': '<s>'})
 
             if not hasattr(self.tokenizer, 'eos_token'):
-                if hasattr(self.tokenizer.tokenizer, 'eos_id'):
+                if hasattr(self.tokenizer.tokenizer, 'eos_id') and self.tokenizer.tokenizer.eos_id() > 0:
                     self.tokenizer.eos_token = self.tokenizer.tokenizer.id_to_piece(self.tokenizer.tokenizer.eos_id())
+                else:
+                    self.tokenizer.add_special_tokens({'eos_token': '<eos>'})
             else:
                 self.tokenizer.add_special_tokens({'eos_token': '</s>'})
 
-            additional_tokens = [f'<extra_id_{i}>' for i in range(self.num_sentinel_tokens)]
-            self.tokenizer.add_special_tokens(additional_tokens)
+            # Special check to see if <extra_id_{}> is already present in the tokenizer. If it is, only modify the additional_special_tokens function.
+            for i in range(self.num_sentinel_tokens):
+                if f'▁<extra_id_{i}>' in self.tokenizer.vocab:
+                    self.tokenizer.special_token_to_id[f'<extra_id_{i}>'] = self.tokenizer.text_to_ids(
+                        f'<extra_id_{i}>'
+                    )[0]
+                else:
+                    self.tokenizer.add_special_tokens([f'<extra_id_{i}>'])
 
     def list_available_models():
         pass
diff --git a/nemo/collections/nlp/modules/common/lm_utils.py b/nemo/collections/nlp/modules/common/lm_utils.py
@@ -87,9 +87,10 @@ def get_lm_model(
         )
 
     if nemo_file is not None:
-        from nemo.collections.nlp.models.language_modeling.megatron_bert_model import MegatronBertModel
         import torch
 
+        from nemo.collections.nlp.models.language_modeling.megatron_bert_model import MegatronBertModel
+
         class Identity(torch.nn.Module):
             def __init__(self):
                 super(Identity, self).__init__()