Fix QA example (#30580)

* Handle cases when CLS token is absent

* Use BOS token as a fallback

examples/pytorch/question-answering/run_qa.py

@@ -434,7 +434,12 @@ def prepare_train_features(examples):
  for i, offsets in enumerate(offset_mapping):
  # We will label impossible answers with the index of the CLS token.
  input_ids = tokenized_examples["input_ids"][i]
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).
  sequence_ids = tokenized_examples.sequence_ids(i)

examples/pytorch/question-answering/run_qa_beam_search.py

@@ -417,7 +417,12 @@ def prepare_train_features(examples):
  for i, offsets in enumerate(offset_mapping):
  # We will label impossible answers with the index of the CLS token.
  input_ids = tokenized_examples["input_ids"][i]
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
  tokenized_examples["cls_index"].append(cls_index)
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).
@@ -534,7 +539,12 @@ def prepare_validation_features(examples):
 
  for i, input_ids in enumerate(tokenized_examples["input_ids"]):
  # Find the CLS token in the input ids.
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
  tokenized_examples["cls_index"].append(cls_index)
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).

examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py

@@ -444,7 +444,12 @@ def prepare_train_features(examples):
  for i, offsets in enumerate(offset_mapping):
  # We will label impossible answers with the index of the CLS token.
  input_ids = tokenized_examples["input_ids"][i]
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
  tokenized_examples["cls_index"].append(cls_index)
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).
@@ -563,7 +568,12 @@ def prepare_validation_features(examples):
 
  for i, input_ids in enumerate(tokenized_examples["input_ids"]):
  # Find the CLS token in the input ids.
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
  tokenized_examples["cls_index"].append(cls_index)
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -513,7 +513,12 @@ def prepare_train_features(examples):
  for i, offsets in enumerate(offset_mapping):
  # We will label impossible answers with the index of the CLS token.
  input_ids = tokenized_examples["input_ids"][i]
- cls_index = input_ids.index(tokenizer.cls_token_id)
+ if tokenizer.cls_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.cls_token_id)
+ elif tokenizer.bos_token_id in input_ids:
+ cls_index = input_ids.index(tokenizer.bos_token_id)
+ else:
+ cls_index = 0
 
  # Grab the sequence corresponding to that example (to know what is the context and what is the question).
  sequence_ids = tokenized_examples.sequence_ids(i)