AL-160: support fast tokenizer

src/transformers/data/processors/squad.py

@@ -8,7 +8,8 @@
 
 from ...file_utils import is_tf_available, is_torch_available
 from ...tokenization_bert import whitespace_tokenize
-from ...tokenization_utils_base import TruncationStrategy
+from ...tokenization_utils_base import PaddingStrategy, TruncationStrategy
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
 from ...utils import logging
 from .utils import DataProcessor
 
@@ -107,6 +108,14 @@ def squad_convert_example_to_features(
     tok_to_orig_index = []
     orig_to_tok_index = []
     all_doc_tokens = []
+    if isinstance(tokenizer, PreTrainedTokenizerFast):
+        tokenizer.set_truncation_and_padding(
+            padding_strategy=PaddingStrategy.DO_NOT_PAD,
+            truncation_strategy=TruncationStrategy.LONGEST_FIRST,
+            max_length=64,
+            stride=0,
+            pad_to_multiple_of=None,
+        )
     for (i, token) in enumerate(example.doc_tokens):
         orig_to_tok_index.append(len(all_doc_tokens))
         sub_tokens = tokenizer.tokenize(token)
@@ -131,6 +140,12 @@ def squad_convert_example_to_features(
         example.question_text, add_special_tokens=False, truncation=True, max_length=max_query_length
     )
 
+    # Handle case where tokenized query is empty, since the fast tokenizer doesn't do so
+    if len(truncated_query) == 0:
+        raise ValueError(
+            f"Input {truncated_query} is not valid. Should be a string or a list/tuple of strings when `is_pretokenized=True`."
+        )
+
     # Tokenizers who insert 2 SEP tokens in-between <context> & <question> need to have special handling
     # in the way they compute mask of added tokens.
     tokenizer_type = type(tokenizer).__name__.replace("Tokenizer", "").lower()
@@ -146,12 +161,28 @@ def squad_convert_example_to_features(
 
         # Define the side we want to truncate / pad and the text/pair sorting
         if tokenizer.padding_side == "right":
-            texts = truncated_query
-            pairs = span_doc_tokens
+            texts = (
+                truncated_query
+                if not isinstance(tokenizer, PreTrainedTokenizerFast)
+                else tokenizer.decode(truncated_query)
+            )
+            # Needed because some tokenizers seem to produce actual tokens,
+            # while others produce token_ids for overflow tokens
+            if isinstance(span_doc_tokens[0], str):
+                pairs = " ".join(span_doc_tokens).replace(" ##", "").strip()
+            else:
+                pairs = span_doc_tokens
             truncation = TruncationStrategy.ONLY_SECOND.value
         else:
-            texts = span_doc_tokens
-            pairs = truncated_query
+            if isinstance(span_doc_tokens[0], str):
+                texts = " ".join(span_doc_tokens).replace(" ##", "").strip()
+            else:
+                texts = span_doc_tokens
+            pairs = (
+                truncated_query
+                if not isinstance(tokenizer, PreTrainedTokenizerFast)
+                else tokenizer.decode(truncated_query)
+            )
             truncation = TruncationStrategy.ONLY_FIRST.value
 
         encoded_dict = tokenizer.encode_plus(  # TODO(thom) update this logic
@@ -165,6 +196,10 @@ def squad_convert_example_to_features(
             return_token_type_ids=True,
         )
 
+        # Handle case where fast tokenizer returns list[list[int]]
+        if isinstance(encoded_dict["input_ids"][0], list):
+            encoded_dict = {k: v[0] for k, v in encoded_dict.items()}
+
         paragraph_len = min(
             len(all_doc_tokens) - len(spans) * doc_stride,
             max_seq_length - len(truncated_query) - sequence_pair_added_tokens,

tests/test_pipelines.py

@@ -672,12 +672,16 @@ def test_torch_question_answering(self):
         for model_name in QA_FINETUNED_MODELS:
             nlp = pipeline(task="question-answering", model=model_name, tokenizer=model_name)
             self._test_qa_pipeline(nlp)
+            nlp = pipeline(task="question-answering", model=model_name, tokenizer=(model_name, {"use_fast": True}))
+            self._test_qa_pipeline(nlp)
 
     @require_tf
     def test_tf_question_answering(self):
         for model_name in QA_FINETUNED_MODELS:
             nlp = pipeline(task="question-answering", model=model_name, tokenizer=model_name, framework="tf")
             self._test_qa_pipeline(nlp)
+            nlp = pipeline(task="question-answering", model=model_name, tokenizer=(model_name, {"use_fast": True}))
+            self._test_qa_pipeline(nlp)
 
 
 class NerPipelineTests(unittest.TestCase):