SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Merge pull request #1984 from huggingface/squad-refactor

Browse Source 
[WIP] Squad refactor
This commit is contained in:
Thomas Wolf
2019-12-10 11:07:26 +01:00
committed by GitHub
parent ecabbf6d28 2a4ef098d6
commit e57d00ee10
15 changed files with 1852 additions and 1448 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
transformers/__init__.py
View File

@@ -26,7 +26,9 @@ from .data import (is_sklearn_available,
InputExample, InputFeatures, DataProcessor,
glue_output_modes, glue_convert_examples_to_features,
glue_processors, glue_tasks_num_labels,
xnli_output_modes, xnli_processors, xnli_tasks_num_labels)
xnli_output_modes, xnli_processors, xnli_tasks_num_labels,
squad_convert_examples_to_features, SquadFeatures,
SquadExample, SquadV1Processor, SquadV2Processor)
if is_sklearn_available():
from .data import glue_compute_metrics, xnli_compute_metrics

3
transformers/data/__init__.py
View File

@@ -1,5 +1,6 @@
from .processors import InputExample, InputFeatures, DataProcessor
from .processors import InputExample, InputFeatures, DataProcessor, SquadFeatures
from .processors import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
from .processors import squad_convert_examples_to_features, SquadExample, SquadV1Processor, SquadV2Processor
from .processors import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
from .metrics import is_sklearn_available

758
transformers/data/metrics/squad_metrics.py Normal file
View File

@@ -0,0 +1,758 @@
""" Very heavily inspired by the official evaluation script for SQuAD version 2.0 which was
modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
In addition to basic functionality, we also compute additional statistics and
plot precision-recall curves if an additional na_prob.json file is provided.
This file is expected to map question ID's to the model's predicted probability
that a question is unanswerable.
"""
import json
import logging
import math
import collections
from io import open
from tqdm import tqdm
import string
import re
from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
logger = logging.getLogger(__name__)
def normalize_answer(s):
"""Lower text and remove punctuation, articles and extra whitespace."""
def remove_articles(text):
regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
return re.sub(regex, ' ', text)
def white_space_fix(text):
return ' '.join(text.split())
def remove_punc(text):
exclude = set(string.punctuation)
return ''.join(ch for ch in text if ch not in exclude)
def lower(text):
return text.lower()
return white_space_fix(remove_articles(remove_punc(lower(s))))
def get_tokens(s):
if not s:
return []
return normalize_answer(s).split()
def compute_exact(a_gold, a_pred):
return int(normalize_answer(a_gold) == normalize_answer(a_pred))
def compute_f1(a_gold, a_pred):
gold_toks = get_tokens(a_gold)
pred_toks = get_tokens(a_pred)
common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
num_same = sum(common.values())
if len(gold_toks) == 0 or len(pred_toks) == 0:
# If either is no-answer, then F1 is 1 if they agree, 0 otherwise
return int(gold_toks == pred_toks)
if num_same == 0:
return 0
precision = 1.0 * num_same / len(pred_toks)
recall = 1.0 * num_same / len(gold_toks)
f1 = (2 * precision * recall) / (precision + recall)
return f1
def get_raw_scores(examples, preds):
"""
Computes the exact and f1 scores from the examples and the model predictions
"""
exact_scores = {}
f1_scores = {}
for example in examples:
qas_id = example.qas_id
gold_answers = [answer['text'] for answer in example.answers if normalize_answer(answer['text'])]
if not gold_answers:
# For unanswerable questions, only correct answer is empty string
gold_answers = ['']
if qas_id not in preds:
print('Missing prediction for %s' % qas_id)
continue
prediction = preds[qas_id]
exact_scores[qas_id] = max(compute_exact(a, prediction) for a in gold_answers)
f1_scores[qas_id] = max(compute_f1(a, prediction) for a in gold_answers)
return exact_scores, f1_scores
def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
new_scores = {}
for qid, s in scores.items():
pred_na = na_probs[qid] > na_prob_thresh
if pred_na:
new_scores[qid] = float(not qid_to_has_ans[qid])
else:
new_scores[qid] = s
return new_scores
def make_eval_dict(exact_scores, f1_scores, qid_list=None):
if not qid_list:
total = len(exact_scores)
return collections.OrderedDict([
('exact', 100.0 * sum(exact_scores.values()) / total),
('f1', 100.0 * sum(f1_scores.values()) / total),
('total', total),
])
else:
total = len(qid_list)
return collections.OrderedDict([
('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
('total', total),
])
def merge_eval(main_eval, new_eval, prefix):
for k in new_eval:
main_eval['%s_%s' % (prefix, k)] = new_eval[k]
def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
cur_score = num_no_ans
best_score = cur_score
best_thresh = 0.0
qid_list = sorted(na_probs, key=lambda k: na_probs[k])
for i, qid in enumerate(qid_list):
if qid not in scores:
continue
if qid_to_has_ans[qid]:
diff = scores[qid]
else:
if preds[qid]:
diff = -1
else:
diff = 0
cur_score += diff
if cur_score > best_score:
best_score = cur_score
best_thresh = na_probs[qid]
has_ans_score, has_ans_cnt = 0, 0
for qid in qid_list:
if not qid_to_has_ans[qid]:
continue
has_ans_cnt += 1
if qid not in scores:
continue
has_ans_score += scores[qid]
return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(
preds, exact_raw, na_probs, qid_to_has_ans)
best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(
preds, f1_raw, na_probs, qid_to_has_ans)
main_eval['best_exact'] = best_exact
main_eval['best_exact_thresh'] = exact_thresh
main_eval['best_f1'] = best_f1
main_eval['best_f1_thresh'] = f1_thresh
main_eval['has_ans_exact'] = has_ans_exact
main_eval['has_ans_f1'] = has_ans_f1
def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
cur_score = num_no_ans
best_score = cur_score
best_thresh = 0.0
qid_list = sorted(na_probs, key=lambda k: na_probs[k])
for _, qid in enumerate(qid_list):
if qid not in scores:
continue
if qid_to_has_ans[qid]:
diff = scores[qid]
else:
if preds[qid]:
diff = -1
else:
diff = 0
cur_score += diff
if cur_score > best_score:
best_score = cur_score
best_thresh = na_probs[qid]
return 100.0 * best_score / len(scores), best_thresh
def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
main_eval['best_exact'] = best_exact
main_eval['best_exact_thresh'] = exact_thresh
main_eval['best_f1'] = best_f1
main_eval['best_f1_thresh'] = f1_thresh
def squad_evaluate(examples, preds, no_answer_probs=None, no_answer_probability_threshold=1.0):
qas_id_to_has_answer = {example.qas_id: bool(example.answers) for example in examples}
has_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if has_answer]
no_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if not has_answer]
if no_answer_probs is None:
no_answer_probs = {k: 0.0 for k in preds}
exact, f1 = get_raw_scores(examples, preds)
exact_threshold = apply_no_ans_threshold(exact, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
f1_threshold = apply_no_ans_threshold(f1, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
evaluation = make_eval_dict(exact_threshold, f1_threshold)
if has_answer_qids:
has_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=has_answer_qids)
merge_eval(evaluation, has_ans_eval, 'HasAns')
if no_answer_qids:
no_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=no_answer_qids)
merge_eval(evaluation, no_ans_eval, 'NoAns')
if no_answer_probs:
find_all_best_thresh(evaluation, preds, exact, f1, no_answer_probs, qas_id_to_has_answer)
return evaluation
def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
"""Project the tokenized prediction back to the original text."""
# When we created the data, we kept track of the alignment between original
# (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
# now `orig_text` contains the span of our original text corresponding to the
# span that we predicted.
#
# However, `orig_text` may contain extra characters that we don't want in
# our prediction.
#
# For example, let's say:
# pred_text = steve smith
# orig_text = Steve Smith's
#
# We don't want to return `orig_text` because it contains the extra "'s".
#
# We don't want to return `pred_text` because it's already been normalized
# (the SQuAD eval script also does punctuation stripping/lower casing but
# our tokenizer does additional normalization like stripping accent
# characters).
#
# What we really want to return is "Steve Smith".
#
# Therefore, we have to apply a semi-complicated alignment heuristic between
# `pred_text` and `orig_text` to get a character-to-character alignment. This
# can fail in certain cases in which case we just return `orig_text`.
def _strip_spaces(text):
ns_chars = []
ns_to_s_map = collections.OrderedDict()
for (i, c) in enumerate(text):
if c == " ":
continue
ns_to_s_map[len(ns_chars)] = i
ns_chars.append(c)
ns_text = "".join(ns_chars)
return (ns_text, ns_to_s_map)
# We first tokenize `orig_text`, strip whitespace from the result
# and `pred_text`, and check if they are the same length. If they are
# NOT the same length, the heuristic has failed. If they are the same
# length, we assume the characters are one-to-one aligned.
tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
tok_text = " ".join(tokenizer.tokenize(orig_text))
start_position = tok_text.find(pred_text)
if start_position == -1:
if verbose_logging:
logger.info(
"Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
return orig_text
end_position = start_position + len(pred_text) - 1
(orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
(tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
if len(orig_ns_text) != len(tok_ns_text):
if verbose_logging:
logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
orig_ns_text, tok_ns_text)
return orig_text
# We then project the characters in `pred_text` back to `orig_text` using
# the character-to-character alignment.
tok_s_to_ns_map = {}
for (i, tok_index) in tok_ns_to_s_map.items():
tok_s_to_ns_map[tok_index] = i
orig_start_position = None
if start_position in tok_s_to_ns_map:
ns_start_position = tok_s_to_ns_map[start_position]
if ns_start_position in orig_ns_to_s_map:
orig_start_position = orig_ns_to_s_map[ns_start_position]
if orig_start_position is None:
if verbose_logging:
logger.info("Couldn't map start position")
return orig_text
orig_end_position = None
if end_position in tok_s_to_ns_map:
ns_end_position = tok_s_to_ns_map[end_position]
if ns_end_position in orig_ns_to_s_map:
orig_end_position = orig_ns_to_s_map[ns_end_position]
if orig_end_position is None:
if verbose_logging:
logger.info("Couldn't map end position")
return orig_text
output_text = orig_text[orig_start_position:(orig_end_position + 1)]
return output_text
def _get_best_indexes(logits, n_best_size):
"""Get the n-best logits from a list."""
index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
best_indexes = []
for i in range(len(index_and_score)):
if i >= n_best_size:
break
best_indexes.append(index_and_score[i][0])
return best_indexes
def _compute_softmax(scores):
"""Compute softmax probability over raw logits."""
if not scores:
return []
max_score = None
for score in scores:
if max_score is None or score > max_score:
max_score = score
exp_scores = []
total_sum = 0.0
for score in scores:
x = math.exp(score - max_score)
exp_scores.append(x)
total_sum += x
probs = []
for score in exp_scores:
probs.append(score / total_sum)
return probs
def compute_predictions_logits(
all_examples,
all_features,
all_results,
n_best_size,
max_answer_length,
do_lower_case,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
verbose_logging,
version_2_with_negative,
null_score_diff_threshold
):
"""Write final predictions to the json file and log-odds of null if needed."""
logger.info("Writing predictions to: %s" % (output_prediction_file))
logger.info("Writing nbest to: %s" % (output_nbest_file))
example_index_to_features = collections.defaultdict(list)
for feature in all_features:
example_index_to_features[feature.example_index].append(feature)
unique_id_to_result = {}
for result in all_results:
unique_id_to_result[result.unique_id] = result
_PrelimPrediction = collections.namedtuple( # pylint: disable=invalid-name
"PrelimPrediction",
["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
all_predictions = collections.OrderedDict()
all_nbest_json = collections.OrderedDict()
scores_diff_json = collections.OrderedDict()
for (example_index, example) in enumerate(all_examples):
features = example_index_to_features[example_index]
prelim_predictions = []
# keep track of the minimum score of null start+end of position 0
score_null = 1000000 # large and positive
min_null_feature_index = 0 # the paragraph slice with min null score
null_start_logit = 0 # the start logit at the slice with min null score
null_end_logit = 0 # the end logit at the slice with min null score
for (feature_index, feature) in enumerate(features):
result = unique_id_to_result[feature.unique_id]
start_indexes = _get_best_indexes(result.start_logits, n_best_size)
end_indexes = _get_best_indexes(result.end_logits, n_best_size)
# if we could have irrelevant answers, get the min score of irrelevant
if version_2_with_negative:
feature_null_score = result.start_logits[0] + result.end_logits[0]
if feature_null_score < score_null:
score_null = feature_null_score
min_null_feature_index = feature_index
null_start_logit = result.start_logits[0]
null_end_logit = result.end_logits[0]
for start_index in start_indexes:
for end_index in end_indexes:
# We could hypothetically create invalid predictions, e.g., predict
# that the start of the span is in the question. We throw out all
# invalid predictions.
if start_index >= len(feature.tokens):
continue
if end_index >= len(feature.tokens):
continue
if start_index not in feature.token_to_orig_map:
continue
if end_index not in feature.token_to_orig_map:
continue
if not feature.token_is_max_context.get(start_index, False):
continue
if end_index < start_index:
continue
length = end_index - start_index + 1
if length > max_answer_length:
continue
prelim_predictions.append(
_PrelimPrediction(
feature_index=feature_index,
start_index=start_index,
end_index=end_index,
start_logit=result.start_logits[start_index],
end_logit=result.end_logits[end_index]))
if version_2_with_negative:
prelim_predictions.append(
_PrelimPrediction(
feature_index=min_null_feature_index,
start_index=0,
end_index=0,
start_logit=null_start_logit,
end_logit=null_end_logit))
prelim_predictions = sorted(
prelim_predictions,
key=lambda x: (x.start_logit + x.end_logit),
reverse=True)
_NbestPrediction = collections.namedtuple( # pylint: disable=invalid-name
"NbestPrediction", ["text", "start_logit", "end_logit"])
seen_predictions = {}
nbest = []
for pred in prelim_predictions:
if len(nbest) >= n_best_size:
break
feature = features[pred.feature_index]
if pred.start_index > 0: # this is a non-null prediction
tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
orig_doc_start = feature.token_to_orig_map[pred.start_index]
orig_doc_end = feature.token_to_orig_map[pred.end_index]
orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
tok_text = " ".join(tok_tokens)
# De-tokenize WordPieces that have been split off.
tok_text = tok_text.replace(" ##", "")
tok_text = tok_text.replace("##", "")
# Clean whitespace
tok_text = tok_text.strip()
tok_text = " ".join(tok_text.split())
orig_text = " ".join(orig_tokens)
final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
if final_text in seen_predictions:
continue
seen_predictions[final_text] = True
else:
final_text = ""
seen_predictions[final_text] = True
nbest.append(
_NbestPrediction(
text=final_text,
start_logit=pred.start_logit,
end_logit=pred.end_logit))
# if we didn't include the empty option in the n-best, include it
if version_2_with_negative:
if "" not in seen_predictions:
nbest.append(
_NbestPrediction(
text="",
start_logit=null_start_logit,
end_logit=null_end_logit))
# In very rare edge cases we could only have single null prediction.
# So we just create a nonce prediction in this case to avoid failure.
if len(nbest) == 1:
nbest.insert(0,
_NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
# In very rare edge cases we could have no valid predictions. So we
# just create a nonce prediction in this case to avoid failure.
if not nbest:
nbest.append(
_NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
assert len(nbest) >= 1
total_scores = []
best_non_null_entry = None
for entry in nbest:
total_scores.append(entry.start_logit + entry.end_logit)
if not best_non_null_entry:
if entry.text:
best_non_null_entry = entry
probs = _compute_softmax(total_scores)
nbest_json = []
for (i, entry) in enumerate(nbest):
output = collections.OrderedDict()
output["text"] = entry.text
output["probability"] = probs[i]
output["start_logit"] = entry.start_logit
output["end_logit"] = entry.end_logit
nbest_json.append(output)
assert len(nbest_json) >= 1
if not version_2_with_negative:
all_predictions[example.qas_id] = nbest_json[0]["text"]
else:
# predict "" iff the null score - the score of best non-null > threshold
score_diff = score_null - best_non_null_entry.start_logit - (
best_non_null_entry.end_logit)
scores_diff_json[example.qas_id] = score_diff
if score_diff > null_score_diff_threshold:
all_predictions[example.qas_id] = ""
else:
all_predictions[example.qas_id] = best_non_null_entry.text
all_nbest_json[example.qas_id] = nbest_json
with open(output_prediction_file, "w") as writer:
writer.write(json.dumps(all_predictions, indent=4) + "\n")
with open(output_nbest_file, "w") as writer:
writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
if version_2_with_negative:
with open(output_null_log_odds_file, "w") as writer:
writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
return all_predictions
def compute_predictions_log_probs(
all_examples,
all_features,
all_results,
n_best_size,
max_answer_length,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
start_n_top,
end_n_top,
version_2_with_negative,
tokenizer,
verbose_logging
):
""" XLNet write prediction logic (more complex than Bert's).
Write final predictions to the json file and log-odds of null if needed.
Requires utils_squad_evaluate.py
"""
_PrelimPrediction = collections.namedtuple( # pylint: disable=invalid-name
"PrelimPrediction",
["feature_index", "start_index", "end_index",
"start_log_prob", "end_log_prob"])
_NbestPrediction = collections.namedtuple( # pylint: disable=invalid-name
"NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
logger.info("Writing predictions to: %s", output_prediction_file)
# logger.info("Writing nbest to: %s" % (output_nbest_file))
example_index_to_features = collections.defaultdict(list)
for feature in all_features:
example_index_to_features[feature.example_index].append(feature)
unique_id_to_result = {}
for result in all_results:
unique_id_to_result[result.unique_id] = result
all_predictions = collections.OrderedDict()
all_nbest_json = collections.OrderedDict()
scores_diff_json = collections.OrderedDict()
for (example_index, example) in enumerate(all_examples):
features = example_index_to_features[example_index]
prelim_predictions = []
# keep track of the minimum score of null start+end of position 0
score_null = 1000000 # large and positive
for (feature_index, feature) in enumerate(features):
result = unique_id_to_result[feature.unique_id]
cur_null_score = result.cls_logits
# if we could have irrelevant answers, get the min score of irrelevant
score_null = min(score_null, cur_null_score)
for i in range(start_n_top):
for j in range(end_n_top):
start_log_prob = result.start_logits[i]
start_index = result.start_top_index[i]
j_index = i * end_n_top + j
end_log_prob = result.end_logits[j_index]
end_index = result.end_top_index[j_index]
# We could hypothetically create invalid predictions, e.g., predict
# that the start of the span is in the question. We throw out all
# invalid predictions.
if start_index >= feature.paragraph_len - 1:
continue
if end_index >= feature.paragraph_len - 1:
continue
if not feature.token_is_max_context.get(start_index, False):
continue
if end_index < start_index:
continue
length = end_index - start_index + 1
if length > max_answer_length:
continue
prelim_predictions.append(
_PrelimPrediction(
feature_index=feature_index,
start_index=start_index,
end_index=end_index,
start_log_prob=start_log_prob,
end_log_prob=end_log_prob))
prelim_predictions = sorted(
prelim_predictions,
key=lambda x: (x.start_log_prob + x.end_log_prob),
reverse=True)
seen_predictions = {}
nbest = []
for pred in prelim_predictions:
if len(nbest) >= n_best_size:
break
feature = features[pred.feature_index]
# XLNet un-tokenizer
# Let's keep it simple for now and see if we need all this later.
#
# tok_start_to_orig_index = feature.tok_start_to_orig_index
# tok_end_to_orig_index = feature.tok_end_to_orig_index
# start_orig_pos = tok_start_to_orig_index[pred.start_index]
# end_orig_pos = tok_end_to_orig_index[pred.end_index]
# paragraph_text = example.paragraph_text
# final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
# Previously used Bert untokenizer
tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
orig_doc_start = feature.token_to_orig_map[pred.start_index]
orig_doc_end = feature.token_to_orig_map[pred.end_index]
orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
# Clean whitespace
tok_text = tok_text.strip()
tok_text = " ".join(tok_text.split())
orig_text = " ".join(orig_tokens)
final_text = get_final_text(tok_text, orig_text, tokenizer.do_lower_case,
verbose_logging)
if final_text in seen_predictions:
continue
seen_predictions[final_text] = True
nbest.append(
_NbestPrediction(
text=final_text,
start_log_prob=pred.start_log_prob,
end_log_prob=pred.end_log_prob))
# In very rare edge cases we could have no valid predictions. So we
# just create a nonce prediction in this case to avoid failure.
if not nbest:
nbest.append(
_NbestPrediction(text="", start_log_prob=-1e6,
end_log_prob=-1e6))
total_scores = []
best_non_null_entry = None
for entry in nbest:
total_scores.append(entry.start_log_prob + entry.end_log_prob)
if not best_non_null_entry:
best_non_null_entry = entry
probs = _compute_softmax(total_scores)
nbest_json = []
for (i, entry) in enumerate(nbest):
output = collections.OrderedDict()
output["text"] = entry.text
output["probability"] = probs[i]
output["start_log_prob"] = entry.start_log_prob
output["end_log_prob"] = entry.end_log_prob
nbest_json.append(output)
assert len(nbest_json) >= 1
assert best_non_null_entry is not None
score_diff = score_null
scores_diff_json[example.qas_id] = score_diff
# note(zhiliny): always predict best_non_null_entry
# and the evaluation script will search for the best threshold
all_predictions[example.qas_id] = best_non_null_entry.text
all_nbest_json[example.qas_id] = nbest_json
with open(output_prediction_file, "w") as writer:
writer.write(json.dumps(all_predictions, indent=4) + "\n")
with open(output_nbest_file, "w") as writer:
writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
if version_2_with_negative:
with open(output_null_log_odds_file, "w") as writer:
writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
return all_predictions

3
transformers/data/processors/__init__.py
View File

@@ -1,3 +1,4 @@
from .utils import InputExample, InputFeatures, DataProcessor
from .glue import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
from .squad import squad_convert_examples_to_features, SquadFeatures, SquadExample, SquadV1Processor, SquadV2Processor
from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

585
transformers/data/processors/squad.py Normal file
View File

@@ -0,0 +1,585 @@
from tqdm import tqdm
import collections
import logging
import os
import json
import numpy as np
from ...tokenization_bert import BasicTokenizer, whitespace_tokenize
from .utils import DataProcessor, InputExample, InputFeatures
from ...file_utils import is_tf_available, is_torch_available
if is_torch_available():
import torch
from torch.utils.data import TensorDataset
if is_tf_available():
import tensorflow as tf
logger = logging.getLogger(__name__)
def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
orig_answer_text):
"""Returns tokenized answer spans that better match the annotated answer."""
tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
for new_start in range(input_start, input_end + 1):
for new_end in range(input_end, new_start - 1, -1):
text_span = " ".join(doc_tokens[new_start:(new_end + 1)])
if text_span == tok_answer_text:
return (new_start, new_end)
return (input_start, input_end)
def _check_is_max_context(doc_spans, cur_span_index, position):
"""Check if this is the 'max context' doc span for the token."""
best_score = None
best_span_index = None
for (span_index, doc_span) in enumerate(doc_spans):
end = doc_span.start + doc_span.length - 1
if position < doc_span.start:
continue
if position > end:
continue
num_left_context = position - doc_span.start
num_right_context = end - position
score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
if best_score is None or score > best_score:
best_score = score
best_span_index = span_index
return cur_span_index == best_span_index
def _new_check_is_max_context(doc_spans, cur_span_index, position):
"""Check if this is the 'max context' doc span for the token."""
# if len(doc_spans) == 1:
# return True
best_score = None
best_span_index = None
for (span_index, doc_span) in enumerate(doc_spans):
end = doc_span["start"] + doc_span["length"] - 1
if position < doc_span["start"]:
continue
if position > end:
continue
num_left_context = position - doc_span["start"]
num_right_context = end - position
score = min(num_left_context, num_right_context) + 0.01 * doc_span["length"]
if best_score is None or score > best_score:
best_score = score
best_span_index = span_index
return cur_span_index == best_span_index
def _is_whitespace(c):
if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
return True
return False
def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
doc_stride, max_query_length, is_training,
return_dataset=False):
"""
Converts a list of examples into a list of features that can be directly given as input to a model.
It is model-dependant and takes advantage of many of the tokenizer's features to create the model's inputs.
Args:
examples: list of :class:`~transformers.data.processors.squad.SquadExample`
tokenizer: an instance of a child of :class:`~transformers.PreTrainedTokenizer`
max_seq_length: The maximum sequence length of the inputs.
doc_stride: The stride used when the context is too large and is split across several features.
max_query_length: The maximum length of the query.
is_training: whether to create features for model evaluation or model training.
return_dataset: Default False. Either 'pt' or 'tf'.
if 'pt': returns a torch.data.TensorDataset,
if 'tf': returns a tf.data.Dataset
Returns:
list of :class:`~transformers.data.processors.squad.SquadFeatures`
Example::
processor = SquadV2Processor()
examples = processor.get_dev_examples(data_dir)
features = squad_convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=not evaluate,
)
"""
# Defining helper methods
unique_id = 1000000000
features = []
for (example_index, example) in enumerate(tqdm(examples)):
if is_training and not example.is_impossible:
# Get start and end position
start_position = example.start_position
end_position = example.end_position
# If the answer cannot be found in the text, then skip this example.
actual_text = " ".join(example.doc_tokens[start_position:(end_position + 1)])
cleaned_answer_text = " ".join(whitespace_tokenize(example.answer_text))
if actual_text.find(cleaned_answer_text) == -1:
logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
continue
tok_to_orig_index = []
orig_to_tok_index = []
all_doc_tokens = []
for (i, token) in enumerate(example.doc_tokens):
orig_to_tok_index.append(len(all_doc_tokens))
sub_tokens = tokenizer.tokenize(token)
for sub_token in sub_tokens:
tok_to_orig_index.append(i)
all_doc_tokens.append(sub_token)
if is_training and not example.is_impossible:
tok_start_position = orig_to_tok_index[example.start_position]
if example.end_position < len(example.doc_tokens) - 1:
tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
else:
tok_end_position = len(all_doc_tokens) - 1
(tok_start_position, tok_end_position) = _improve_answer_span(
all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.answer_text
)
spans = []
truncated_query = tokenizer.encode(example.question_text, add_special_tokens=False, max_length=max_query_length)
sequence_added_tokens = tokenizer.max_len - tokenizer.max_len_single_sentence
sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair
span_doc_tokens = all_doc_tokens
while len(spans) * doc_stride < len(all_doc_tokens):
encoded_dict = tokenizer.encode_plus(
truncated_query if tokenizer.padding_side == "right" else span_doc_tokens,
span_doc_tokens if tokenizer.padding_side == "right" else truncated_query,
max_length=max_seq_length,
return_overflowing_tokens=True,
pad_to_max_length=True,
stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
truncation_strategy='only_second' if tokenizer.padding_side == "right" else 'only_first'
)
paragraph_len = min(len(all_doc_tokens) - len(spans) * doc_stride, max_seq_length - len(truncated_query) - sequence_pair_added_tokens)
if tokenizer.pad_token_id in encoded_dict['input_ids']:
non_padded_ids = encoded_dict['input_ids'][:encoded_dict['input_ids'].index(tokenizer.pad_token_id)]
else:
non_padded_ids = encoded_dict['input_ids']
tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)
token_to_orig_map = {}
for i in range(paragraph_len):
index = len(truncated_query) + sequence_added_tokens + i if tokenizer.padding_side == "right" else i
token_to_orig_map[index] = tok_to_orig_index[len(spans) * doc_stride + i]
encoded_dict["paragraph_len"] = paragraph_len
encoded_dict["tokens"] = tokens
encoded_dict["token_to_orig_map"] = token_to_orig_map
encoded_dict["truncated_query_with_special_tokens_length"] = len(truncated_query) + sequence_added_tokens
encoded_dict["token_is_max_context"] = {}
encoded_dict["start"] = len(spans) * doc_stride
encoded_dict["length"] = paragraph_len
spans.append(encoded_dict)
if "overflowing_tokens" not in encoded_dict:
break
span_doc_tokens = encoded_dict["overflowing_tokens"]
for doc_span_index in range(len(spans)):
for j in range(spans[doc_span_index]["paragraph_len"]):
is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
index = j if tokenizer.padding_side == "left" else spans[doc_span_index]["truncated_query_with_special_tokens_length"] + j
spans[doc_span_index]["token_is_max_context"][index] = is_max_context
for span in spans:
# Identify the position of the CLS token
cls_index = span['input_ids'].index(tokenizer.cls_token_id)
# p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
# Original TF implem also keep the classification token (set to 0) (not sure why...)
p_mask = np.array(span['token_type_ids'])
p_mask = np.minimum(p_mask, 1)
if tokenizer.padding_side == "right":
# Limit positive values to one
p_mask = 1 - p_mask
p_mask[np.where(np.array(span["input_ids"]) == tokenizer.sep_token_id)[0]] = 1
# Set the CLS index to '0'
p_mask[cls_index] = 0
span_is_impossible = example.is_impossible
start_position = 0
end_position = 0
if is_training and not span_is_impossible:
# For training, if our document chunk does not contain an annotation
# we throw it out, since there is nothing to predict.
doc_start = span["start"]
doc_end = span["start"] + span["length"] - 1
out_of_span = False
if not (tok_start_position >= doc_start and tok_end_position <= doc_end):
out_of_span = True
if out_of_span:
start_position = cls_index
end_position = cls_index
span_is_impossible = True
else:
if tokenizer.padding_side == "left":
doc_offset = 0
else:
doc_offset = len(truncated_query) + sequence_added_tokens
start_position = tok_start_position - doc_start + doc_offset
end_position = tok_end_position - doc_start + doc_offset
features.append(SquadFeatures(
span['input_ids'],
span['attention_mask'],
span['token_type_ids'],
cls_index,
p_mask.tolist(),
example_index=example_index,
unique_id=unique_id,
paragraph_len=span['paragraph_len'],
token_is_max_context=span["token_is_max_context"],
tokens=span["tokens"],
token_to_orig_map=span["token_to_orig_map"],
start_position=start_position,
end_position=end_position
))
unique_id += 1
if return_dataset == 'pt':
if not is_torch_available():
raise ImportError("Pytorch must be installed to return a pytorch dataset.")
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_segment_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
if not is_training:
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_example_index, all_cls_index, all_p_mask)
else:
all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_start_positions, all_end_positions,
all_cls_index, all_p_mask)
return features, dataset
return features
class SquadProcessor(DataProcessor):
"""
Processor for the SQuAD data set.
Overriden by SquadV1Processor and SquadV2Processor, used by the version 1.1 and version 2.0 of SQuAD, respectively.
"""
train_file = None
dev_file = None
def _get_example_from_tensor_dict(self, tensor_dict, evaluate=False):
if not evaluate:
answer = tensor_dict['answers']['text'][0].numpy().decode('utf-8')
answer_start = tensor_dict['answers']['answer_start'][0].numpy()
answers = []
else:
answers = [{
"answer_start": start.numpy(),
"text": text.numpy().decode('utf-8')
} for start, text in zip(tensor_dict['answers']["answer_start"], tensor_dict['answers']["text"])]
answer = None
answer_start = None
return SquadExample(
qas_id=tensor_dict['id'].numpy().decode("utf-8"),
question_text=tensor_dict['question'].numpy().decode('utf-8'),
context_text=tensor_dict['context'].numpy().decode('utf-8'),
answer_text=answer,
start_position_character=answer_start,
title=tensor_dict['title'].numpy().decode('utf-8'),
answers=answers
)
def get_examples_from_dataset(self, dataset, evaluate=False):
"""
Creates a list of :class:`~transformers.data.processors.squad.SquadExample` using a TFDS dataset.
Args:
dataset: The tfds dataset loaded from `tensorflow_datasets.load("squad")`
evaluate: boolean specifying if in evaluation mode or in training mode
Returns:
List of SquadExample
Examples::
import tensorflow_datasets as tfds
dataset = tfds.load("squad")
training_examples = get_examples_from_dataset(dataset, evaluate=False)
evaluation_examples = get_examples_from_dataset(dataset, evaluate=True)
"""
if evaluate:
dataset = dataset["validation"]
else:
dataset = dataset["train"]
examples = []
for tensor_dict in tqdm(dataset):
examples.append(self._get_example_from_tensor_dict(tensor_dict, evaluate=evaluate))
return examples
def get_train_examples(self, data_dir, filename=None):
"""
Returns the training examples from the data directory.
Args:
data_dir: Directory containing the data files used for training and evaluating.
filename: None by default, specify this if the training file has a different name than the original one
which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
"""
if self.train_file is None:
raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
with open(os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding='utf-8') as reader:
input_data = json.load(reader)["data"]
return self._create_examples(input_data, "train")
def get_dev_examples(self, data_dir, filename=None):
"""
Returns the evaluation example from the data directory.
Args:
data_dir: Directory containing the data files used for training and evaluating.
filename: None by default, specify this if the evaluation file has a different name than the original one
which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
"""
if self.dev_file is None:
raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
with open(os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding='utf-8') as reader:
input_data = json.load(reader)["data"]
return self._create_examples(input_data, "dev")
def _create_examples(self, input_data, set_type):
is_training = set_type == "train"
examples = []
for entry in tqdm(input_data):
title = entry['title']
for paragraph in entry["paragraphs"]:
context_text = paragraph["context"]
for qa in paragraph["qas"]:
qas_id = qa["id"]
question_text = qa["question"]
start_position_character = None
answer_text = None
answers = []
if "is_impossible" in qa:
is_impossible = qa["is_impossible"]
else:
is_impossible = False
if not is_impossible:
if is_training:
answer = qa["answers"][0]
answer_text = answer['text']
start_position_character = answer['answer_start']
else:
answers = qa["answers"]
example = SquadExample(
qas_id=qas_id,
question_text=question_text,
context_text=context_text,
answer_text=answer_text,
start_position_character=start_position_character,
title=title,
is_impossible=is_impossible,
answers=answers
)
examples.append(example)
return examples
class SquadV1Processor(SquadProcessor):
train_file = "train-v1.1.json"
dev_file = "dev-v1.1.json"
class SquadV2Processor(SquadProcessor):
train_file = "train-v2.0.json"
dev_file = "dev-v2.0.json"
class SquadExample(object):
"""
A single training/test example for the Squad dataset, as loaded from disk.
Args:
qas_id: The example's unique identifier
question_text: The question string
context_text: The context string
answer_text: The answer string
start_position_character: The character position of the start of the answer
title: The title of the example
answers: None by default, this is used during evaluation. Holds answers as well as their start positions.
is_impossible: False by default, set to True if the example has no possible answer.
"""
def __init__(self,
qas_id,
question_text,
context_text,
answer_text,
start_position_character,
title,
answers=[],
is_impossible=False):
self.qas_id = qas_id
self.question_text = question_text
self.context_text = context_text
self.answer_text = answer_text
self.title = title
self.is_impossible = is_impossible
self.answers = answers
self.start_position, self.end_position = 0, 0
doc_tokens = []
char_to_word_offset = []
prev_is_whitespace = True
# Split on whitespace so that different tokens may be attributed to their original position.
for c in self.context_text:
if _is_whitespace(c):
prev_is_whitespace = True
else:
if prev_is_whitespace:
doc_tokens.append(c)
else:
doc_tokens[-1] += c
prev_is_whitespace = False
char_to_word_offset.append(len(doc_tokens) - 1)
self.doc_tokens = doc_tokens
self.char_to_word_offset = char_to_word_offset
# Start end end positions only has a value during evaluation.
if start_position_character is not None and not is_impossible:
self.start_position = char_to_word_offset[start_position_character]
self.end_position = char_to_word_offset[start_position_character + len(answer_text) - 1]
class SquadFeatures(object):
"""
Single squad example features to be fed to a model.
Those features are model-specific and can be crafted from :class:`~transformers.data.processors.squad.SquadExample`
using the :method:`~transformers.data.processors.squad.squad_convert_examples_to_features` method.
Args:
input_ids: Indices of input sequence tokens in the vocabulary.
attention_mask: Mask to avoid performing attention on padding token indices.
token_type_ids: Segment token indices to indicate first and second portions of the inputs.
cls_index: the index of the CLS token.
p_mask: Mask identifying tokens that can be answers vs. tokens that cannot.
Mask with 1 for tokens than cannot be in the answer and 0 for token that can be in an answer
example_index: the index of the example
unique_id: The unique Feature identifier
paragraph_len: The length of the context
token_is_max_context: List of booleans identifying which tokens have their maximum context in this feature object.
If a token does not have their maximum context in this feature object, it means that another feature object
has more information related to that token and should be prioritized over this feature for that token.
tokens: list of tokens corresponding to the input ids
token_to_orig_map: mapping between the tokens and the original text, needed in order to identify the answer.
start_position: start of the answer token index
end_position: end of the answer token index
"""
def __init__(self,
input_ids,
attention_mask,
token_type_ids,
cls_index,
p_mask,
example_index,
unique_id,
paragraph_len,
token_is_max_context,
tokens,
token_to_orig_map,
start_position,
end_position
):
self.input_ids = input_ids
self.attention_mask = attention_mask
self.token_type_ids = token_type_ids
self.cls_index = cls_index
self.p_mask = p_mask
self.example_index = example_index
self.unique_id = unique_id
self.paragraph_len = paragraph_len
self.token_is_max_context = token_is_max_context
self.tokens = tokens
self.token_to_orig_map = token_to_orig_map
self.start_position = start_position
self.end_position = end_position
class SquadResult(object):
"""
Constructs a SquadResult which can be used to evaluate a model's output on the SQuAD dataset.
Args:
unique_id: The unique identifier corresponding to that example.
start_logits: The logits corresponding to the start of the answer
end_logits: The logits corresponding to the end of the answer
"""
def __init__(self, unique_id, start_logits, end_logits, start_top_index=None, end_top_index=None, cls_logits=None):
self.start_logits = start_logits
self.end_logits = end_logits
self.unique_id = unique_id
if start_top_index:
self.start_top_index = start_top_index
self.end_top_index = end_top_index
self.cls_logits = cls_logits

87
transformers/tests/tokenization_tests_commons.py
View File

@@ -378,3 +378,90 @@ class CommonTestCases:
special_tokens_mask = tokenizer.get_special_tokens_mask(encoded_sequence_w_special, already_has_special_tokens=True)
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
self.assertEqual(special_tokens_mask_orig, special_tokens_mask)
def test_padding_to_max_length(self):
tokenizer = self.get_tokenizer()
sequence = "Sequence"
padding_size = 10
padding_idx = tokenizer.pad_token_id
# RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
tokenizer.padding_side = "right"
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
padded_sequence = tokenizer.encode(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True)
padded_sequence_length = len(padded_sequence)
assert sequence_length + padding_size == padded_sequence_length
assert encoded_sequence + [padding_idx] * padding_size == padded_sequence
# LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
tokenizer.padding_side = "left"
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
padded_sequence = tokenizer.encode(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True)
padded_sequence_length = len(padded_sequence)
assert sequence_length + padding_size == padded_sequence_length
assert [padding_idx] * padding_size + encoded_sequence == padded_sequence
# RIGHT & LEFT PADDING - Check that nothing is done when a maximum length is not specified
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
tokenizer.padding_side = "right"
padded_sequence_right = tokenizer.encode(sequence, pad_to_max_length=True)
padded_sequence_right_length = len(padded_sequence_right)
tokenizer.padding_side = "left"
padded_sequence_left = tokenizer.encode(sequence, pad_to_max_length=True)
padded_sequence_left_length = len(padded_sequence_left)
assert sequence_length == padded_sequence_right_length
assert encoded_sequence == padded_sequence_right
assert sequence_length == padded_sequence_left_length
assert encoded_sequence == padded_sequence_left
def test_encode_plus_with_padding(self):
tokenizer = self.get_tokenizer()
sequence = "Sequence"
padding_size = 10
padding_idx = tokenizer.pad_token_id
token_type_padding_idx = tokenizer.pad_token_type_id
encoded_sequence = tokenizer.encode_plus(sequence, return_special_tokens_mask=True)
input_ids = encoded_sequence['input_ids']
token_type_ids = encoded_sequence['token_type_ids']
attention_mask = encoded_sequence['attention_mask']
special_tokens_mask = encoded_sequence['special_tokens_mask']
sequence_length = len(input_ids)
# Test right padding
tokenizer.padding_side = "right"
padded_sequence = tokenizer.encode_plus(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True, return_special_tokens_mask=True)
padded_input_ids = padded_sequence['input_ids']
padded_token_type_ids = padded_sequence['token_type_ids']
padded_attention_mask = padded_sequence['attention_mask']
padded_special_tokens_mask = padded_sequence['special_tokens_mask']
padded_sequence_length = len(padded_input_ids)
assert sequence_length + padding_size == padded_sequence_length
assert input_ids + [padding_idx] * padding_size == padded_input_ids
assert token_type_ids + [token_type_padding_idx] * padding_size == padded_token_type_ids
assert attention_mask + [0] * padding_size == padded_attention_mask
assert special_tokens_mask + [1] * padding_size == padded_special_tokens_mask
# Test left padding
tokenizer.padding_side = "left"
padded_sequence = tokenizer.encode_plus(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True, return_special_tokens_mask=True)
padded_input_ids = padded_sequence['input_ids']
padded_token_type_ids = padded_sequence['token_type_ids']
padded_attention_mask = padded_sequence['attention_mask']
padded_special_tokens_mask = padded_sequence['special_tokens_mask']
padded_sequence_length = len(padded_input_ids)
assert sequence_length + padding_size == padded_sequence_length
assert [padding_idx] * padding_size + input_ids == padded_input_ids
assert [token_type_padding_idx] * padding_size + token_type_ids == padded_token_type_ids
assert [0] * padding_size + attention_mask == padded_attention_mask
assert [1] * padding_size + special_tokens_mask == padded_special_tokens_mask

82
transformers/tokenization_utils.py
View File

@@ -78,6 +78,8 @@ class PreTrainedTokenizer(object):
"pad_token", "cls_token", "mask_token",
"additional_special_tokens"]
padding_side = "right"
@property
def bos_token(self):
""" Beginning of sentence token (string). Log an error if used while not having been set. """
@@ -191,6 +193,11 @@ class PreTrainedTokenizer(object):
""" Id of the padding token in the vocabulary. Log an error if used while not having been set. """
return self.convert_tokens_to_ids(self.pad_token)
@property
def pad_token_type_id(self):
""" Id of the padding token type in the vocabulary."""
return self._pad_token_type_id
@property
def cls_token_id(self):
""" Id of the classification token in the vocabulary. E.g. to extract a summary of an input sequence leveraging self-attention along the full depth of the model. Log an error if used while not having been set. """
@@ -214,10 +221,14 @@ class PreTrainedTokenizer(object):
self._pad_token = None
self._cls_token = None
self._mask_token = None
self._pad_token_type_id = 0
self._additional_special_tokens = []
self.max_len = max_len if max_len is not None else int(1e12)
# Padding side is right by default and over-riden in subclsses. If specified in the kwargs, it is changed.
self.padding_side = kwargs.pop('padding_side', self.padding_side)
# Added tokens
self.added_tokens_encoder = {}
self.added_tokens_decoder = {}
@@ -615,6 +626,10 @@ class PreTrainedTokenizer(object):
vocabularies (BPE/SentencePieces/WordPieces).
Take care of added tokens.
text: The sequence to be encoded.
return_tokens_mapped_to_origin: (optional) Set to True to return the index of each token in the initial whitespace tokenization. (default False).
**kwargs: passed to the child `self.tokenize()` method
"""
def lowercase_text(t):
# convert non-special tokens to lowercase
@@ -715,6 +730,7 @@ class PreTrainedTokenizer(object):
max_length=None,
stride=0,
truncation_strategy='longest_first',
pad_to_max_length=False,
return_tensors=None,
**kwargs):
"""
@@ -741,6 +757,12 @@ class PreTrainedTokenizer(object):
- 'only_first': Only truncate the first sequence
- 'only_second': Only truncate the second sequence
- 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
The tokenizer padding sides are handled by the following strings:
- 'left': pads on the left of the sequences
- 'right': pads on the right of the sequences
Defaults to False: no padding.
return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
or PyTorch torch.Tensor instead of a list of python integers.
**kwargs: passed to the `self.tokenize()` method
@@ -751,6 +773,7 @@ class PreTrainedTokenizer(object):
add_special_tokens=add_special_tokens,
stride=stride,
truncation_strategy=truncation_strategy,
pad_to_max_length=pad_to_max_length,
return_tensors=return_tensors,
**kwargs)
@@ -763,8 +786,10 @@ class PreTrainedTokenizer(object):
max_length=None,
stride=0,
truncation_strategy='longest_first',
pad_to_max_length=False,
return_tensors=None,
return_token_type_ids=True,
return_attention_mask=True,
return_overflowing_tokens=False,
return_special_tokens_mask=False,
**kwargs):
@@ -791,9 +816,16 @@ class PreTrainedTokenizer(object):
- 'only_first': Only truncate the first sequence
- 'only_second': Only truncate the second sequence
- 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
The tokenizer padding sides are handled by the following strings:
- 'left': pads on the left of the sequences
- 'right': pads on the right of the sequences
Defaults to False: no padding.
return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
or PyTorch torch.Tensor instead of a list of python integers.
return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (default True).
return_attention_mask: (optional) Set to False to avoir returning attention mask (default True)
return_overflowing_tokens: (optional) Set to True to return overflowing token information (default False).
return_special_tokens_mask: (optional) Set to True to return special tokens mask information (default False).
**kwargs: passed to the `self.tokenize()` method
@@ -804,6 +836,7 @@ class PreTrainedTokenizer(object):
{
input_ids: list[int],
token_type_ids: list[int] if return_token_type_ids is True (default)
attention_mask: list[int] if return_attention_mask is True (default)
overflowing_tokens: list[int] if a ``max_length`` is specified and return_overflowing_tokens is True
num_truncated_tokens: int if a ``max_length`` is specified and return_overflowing_tokens is True
special_tokens_mask: list[int] if ``add_special_tokens`` if set to ``True`` and return_special_tokens_mask is True
@@ -812,7 +845,7 @@ class PreTrainedTokenizer(object):
With the fields:
``input_ids``: list of token ids to be fed to a model
``token_type_ids``: list of token type ids to be fed to a model
``attention_mask``: list of indices specifying which tokens should be attended to by the model
``overflowing_tokens``: list of overflowing tokens if a max length is specified.
``num_truncated_tokens``: number of overflowing tokens a ``max_length`` is specified
``special_tokens_mask``: if adding special tokens, this is a list of [0, 1], with 0 specifying special added
@@ -835,18 +868,22 @@ class PreTrainedTokenizer(object):
return self.prepare_for_model(first_ids,
pair_ids=second_ids,
max_length=max_length,
pad_to_max_length=pad_to_max_length,
add_special_tokens=add_special_tokens,
stride=stride,
truncation_strategy=truncation_strategy,
return_tensors=return_tensors,
return_attention_mask=return_attention_mask,
return_token_type_ids=return_token_type_ids,
return_overflowing_tokens=return_overflowing_tokens,
return_special_tokens_mask=return_special_tokens_mask)
def prepare_for_model(self, ids, pair_ids=None, max_length=None, add_special_tokens=True, stride=0,
truncation_strategy='longest_first',
pad_to_max_length=False,
return_tensors=None,
return_token_type_ids=True,
return_attention_mask=True,
return_overflowing_tokens=False,
return_special_tokens_mask=False):
"""
@@ -871,9 +908,16 @@ class PreTrainedTokenizer(object):
- 'only_first': Only truncate the first sequence
- 'only_second': Only truncate the second sequence
- 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
The tokenizer padding sides are handled by the following strings:
- 'left': pads on the left of the sequences
- 'right': pads on the right of the sequences
Defaults to False: no padding.
return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
or PyTorch torch.Tensor instead of a list of python integers.
return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (default True).
return_attention_mask: (optional) Set to False to avoir returning attention mask (default True)
return_overflowing_tokens: (optional) Set to True to return overflowing token information (default False).
return_special_tokens_mask: (optional) Set to True to return special tokens mask information (default False).
@@ -951,7 +995,43 @@ class PreTrainedTokenizer(object):
logger.warning("Token indices sequence length is longer than the specified maximum sequence length "
"for this model ({} > {}). Running this sequence through the model will result in "
"indexing errors".format(len(ids), self.max_len))
needs_to_be_padded = pad_to_max_length and (
max_length and len(encoded_inputs["input_ids"]) < max_length
or
max_length is None and len(encoded_inputs["input_ids"]) < self.max_len and self.max_len <= 10000
)
if pad_to_max_length and max_length is None and self.max_len > 10000:
logger.warning("Sequence can't be padded as the maximum ")
if needs_to_be_padded:
difference = (max_length if max_length is not None else self.max_len) - len(encoded_inputs["input_ids"])
if self.padding_side == 'right':
if return_attention_mask:
encoded_inputs["attention_mask"] = [1] * len(encoded_inputs["input_ids"]) + [0] * difference
if return_token_type_ids:
encoded_inputs["token_type_ids"] = encoded_inputs["token_type_ids"] + [self.pad_token_type_id] * difference
if return_special_tokens_mask:
encoded_inputs["special_tokens_mask"] = encoded_inputs["special_tokens_mask"] + [1] * difference
encoded_inputs["input_ids"] = encoded_inputs["input_ids"] + [self.pad_token_id] * difference
elif self.padding_side == 'left':
if return_attention_mask:
encoded_inputs["attention_mask"] = [0] * difference + [1] * len(encoded_inputs["input_ids"])
if return_token_type_ids:
encoded_inputs["token_type_ids"] = [self.pad_token_type_id] * difference + encoded_inputs["token_type_ids"]
if return_special_tokens_mask:
encoded_inputs["special_tokens_mask"] = [1] * difference + encoded_inputs["special_tokens_mask"]
encoded_inputs["input_ids"] = [self.pad_token_id] * difference + encoded_inputs["input_ids"]
else:
raise ValueError("Invalid padding strategy:" + str(self.padding_side))
elif return_attention_mask:
encoded_inputs["attention_mask"] = [1] * len(encoded_inputs["input_ids"])
return encoded_inputs
def truncate_sequences(self, ids, pair_ids=None, num_tokens_to_remove=0, truncation_strategy='longest_first', stride=0):

2
transformers/tokenization_xlnet.py
View File

@@ -60,6 +60,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
padding_side = "left"
def __init__(self, vocab_file,
do_lower_case=False, remove_space=True, keep_accents=False,
@@ -74,6 +75,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
self.max_len_single_sentence = self.max_len - 2 # take into account special tokens
self.max_len_sentences_pair = self.max_len - 3 # take into account special tokens
self._pad_token_type_id = 3
try:
import sentencepiece as spm