def compute_predictions_logits(
all_examples,
all_features,
all_results,
n_best_size,
max_answer_length,
tokenizer,
output_prediction_file=None,
output_nbest_file=None,
):
"""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
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)
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],
)
)
# start logit + end logit 큰 순서대로 (확률을 합한 값이 높은 순으로)
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","orig_doc_start","orig_doc_end"]
)
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)]
tok_start_index = feature.token_to_orig_map[pred.start_index]
tok_end_index = feature.token_to_orig_map[pred.end_index]
orig_tokens = example.doc_tokens[tok_start_index : (tok_end_index + 1)]
# For High-lighting
orig_doc_start = sum(list(map(lambda x: len(x)+1,example.doc_tokens[0:tok_start_index])))
orig_doc_end = orig_doc_start + sum(list(map(lambda x: len(x),orig_tokens)))
tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
orig_text = " ".join(orig_tokens)
final_text = get_final_text(tok_text,orig_text,do_lower_case=False)
# final_text = tok_text
if final_text in seen_predictions:
continue
seen_predictions[final_text] = True
else:
final_text = ""
seen_predictions[final_text] = True
orig_doc_start=-1
orig_doc_end=-1
nbest.append(_NbestPrediction(text=final_text, start_logit=pred.start_logit, end_logit=pred.end_logit, orig_doc_start=orig_doc_start,orig_doc_end=orig_doc_end))
# 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=-10.0, end_logit=-10.0,orig_doc_start=-1,orig_doc_end=-1))
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
output["doc_start"] = entry.orig_doc_start
output["doc_end"] = entry.orig_doc_end
nbest_json.append(output)
all_nbest_json[example.qas_id] = nbest_json
assert len(nbest_json) >= 1
all_predictions[example.qas_id] = nbest_json[0]["text"]
if not (output_prediction_file == None or output_nbest_file==None):
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")
return all_predictions, all_nbest_json
def compute_predictions_logits_all(
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,
tokenizer,
):
"""This is a function from the transformer library modified to work on the multi-passage setting"""
"""Write final predictions to the json file and log-odds of null if needed."""
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", [
"example_index", "feature_index", "start_index", "end_index",
"start_logit", "end_logit"
])
all_predictions = collections.OrderedDict()
all_nbest_json = collections.OrderedDict()
scores_diff_json = collections.OrderedDict()
prelim_predictions = []
for (example_index, example) in enumerate(all_examples):
features = example_index_to_features[example_index]
# 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(
example_index=example_index,
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
example = all_examples[pred.example_index]
features = example_index_to_features[pred.example_index]
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 = tokenizer.convert_tokens_to_string(tok_tokens)
# 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["0"] = nbest_json[0][
"text"] # all_predictions[example.qas_id] = nbest_json[0]["text"] same below..
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["0"] = ""
else:
all_predictions[example.qas_id] = best_non_null_entry.text
all_nbest_json["0"] = 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, prelim_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)
if hasattr(tokenizer, "do_lower_case"):
do_lower_case = tokenizer.do_lower_case
else:
do_lower_case = tokenizer.do_lowercase_and_remove_accent
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
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