riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-02 03:01:07 +06:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

Update pipelines.py

Browse Source 
Modified QA pipeline to consider all features for each example before generating topk answers. 
Current pipeline only takes one SquadExample, one SquadFeature, one start logit list, one end logit list to retrieve the answer, this is not correct as one SquadExample can produce multiple SquadFeatures.
This commit is contained in:
Rishabh Manoj 2020-01-08 21:12:34 +05:30 committed by GitHub
parent 16ce15ed4b
commit f26a353057
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 53 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

96
src/transformers/pipelines.py
View File

@ -705,55 +705,65 @@ class QuestionAnsweringPipeline(Pipeline):
# Convert inputs to features # Convert inputs to features
examples = self._args_parser(*texts, **kwargs) examples = self._args_parser(*texts, **kwargs)
features = squad_convert_examples_to_features( features_list = [ squad_convert_examples_to_features(
examples, self.tokenizer, kwargs["max_seq_len"], kwargs["doc_stride"], kwargs["max_question_len"], False [example],
) self.tokenizer,
fw_args = self.inputs_for_model([f.__dict__ for f in features]) kwargs["max_seq_len"],
kwargs["doc_stride"],
kwargs["max_question_len"],
False
) for example in examples ]
all_answers = []
for features, example in zip(features_list, examples):
fw_args = self.inputs_for_model([f.__dict__ for f in features])
# Manage tensor allocation on correct device # Manage tensor allocation on correct device
with self.device_placement(): with self.device_placement():
if self.framework == "tf": if self.framework == "tf":
fw_args = {k: tf.constant(v) for (k, v) in fw_args.items()} fw_args = {k: tf.constant(v) for (k, v) in fw_args.items()}
start, end = self.model(fw_args) start, end = self.model(fw_args)
start, end = start.numpy(), end.numpy() start, end = start.numpy(), end.numpy()
else: else:
with torch.no_grad(): with torch.no_grad():
# Retrieve the score for the context tokens only (removing question tokens) # Retrieve the score for the context tokens only (removing question tokens)
fw_args = {k: torch.tensor(v, device=self.device) for (k, v) in fw_args.items()} fw_args = {k: torch.tensor(v, device=self.device) for (k, v) in fw_args.items()}
start, end = self.model(**fw_args) start, end = self.model(**fw_args)
start, end = start.cpu().numpy(), end.cpu().numpy() start, end = start.cpu().numpy(), end.cpu().numpy()
answers = [] answers = []
for (example, feature, start_, end_) in zip(examples, features, start, end): for (feature, start_, end_) in zip(features, start, end):
# Normalize logits and spans to retrieve the answer # Normalize logits and spans to retrieve the answer
start_ = np.exp(start_) / np.sum(np.exp(start_)) start_ = np.exp(start_) / np.sum(np.exp(start_))
end_ = np.exp(end_) / np.sum(np.exp(end_)) end_ = np.exp(end_) / np.sum(np.exp(end_))
# Mask padding and question # Mask padding and question
start_, end_ = start_ * np.abs(np.array(feature.p_mask) - 1), end_ * np.abs(np.array(feature.p_mask) - 1) start_, end_ = start_ * np.abs(np.array(feature.p_mask) - 1), end_ * np.abs(np.array(feature.p_mask) - 1)
# TODO : What happens if not possible # TODO : What happens if not possible
# Mask CLS # Mask CLS
start_[0] = end_[0] = 0 start_[0] = end_[0] = 0
starts, ends, scores = self.decode(start_, end_, kwargs["topk"], kwargs["max_answer_len"]) starts, ends, scores = self.decode(start_, end_, kwargs["topk"], kwargs["max_answer_len"])
char_to_word = np.array(example.char_to_word_offset) char_to_word = np.array(example.char_to_word_offset)
# Convert the answer (tokens) back to the original text # Convert the answer (tokens) back to the original text
answers += [ answers += [
{ {
"score": score.item(), "score": score.item(),
"start": np.where(char_to_word == feature.token_to_orig_map[s])[0][0].item(), "start": np.where(char_to_word == feature.token_to_orig_map[s])[0][0].item(),
"end": np.where(char_to_word == feature.token_to_orig_map[e])[0][-1].item(), "end": np.where(char_to_word == feature.token_to_orig_map[e])[0][-1].item(),
"answer": " ".join( "answer": " ".join(
example.doc_tokens[feature.token_to_orig_map[s] : feature.token_to_orig_map[e] + 1] example.doc_tokens[feature.token_to_orig_map[s] : feature.token_to_orig_map[e] + 1]
), ),
} }
for s, e, score in zip(starts, ends, scores) for s, e, score in zip(starts, ends, scores)
] ]
if len(answers) == 1: answers = sorted(answers, key = lambda x:x['score'], reverse=True)[:kwargs["topk"]]
return answers[0] all_answers+=answers
return answers
if len(all_answers) == 1:
return all_answers[0]
return all_answers
def decode(self, start: np.ndarray, end: np.ndarray, topk: int, max_answer_len: int) -> Tuple: def decode(self, start: np.ndarray, end: np.ndarray, topk: int, max_answer_len: int) -> Tuple:
""" """