def reduce_aggregated_logs(self, aggregated_logs):
all_predictions, _, scores_diff = (self.squad_lib.postprocess_output(
self._eval_examples,
self._eval_features,
aggregated_logs,
self.task_config.n_best_size,
self.task_config.max_answer_length,
self.task_config.validation_data.do_lower_case,
version_2_with_negative=(
self.task_config.validation_data.version_2_with_negative),
null_score_diff_threshold=(
self.task_config.null_score_diff_threshold),
verbose=False))
with tf.io.gfile.GFile(self.task_config.validation_data.input_path,
'r') as reader:
dataset_json = json.load(reader)
pred_dataset = dataset_json['data']
if self.task_config.validation_data.version_2_with_negative:
eval_metrics = squad_evaluate_v2_0.evaluate(
pred_dataset, all_predictions, scores_diff)
else:
eval_metrics = squad_evaluate_v1_1.evaluate(
pred_dataset, all_predictions)
return eval_metrics
def eval_squad(strategy,
input_meta_data,
tokenizer,
bert_config,
squad_lib,
init_checkpoint=None):
"""Get prediction results and evaluate them against ground truth."""
if init_checkpoint is None:
init_checkpoint = tf.train.latest_checkpoint(FLAGS.model_dir)
all_predict_files = _get_matched_files(FLAGS.predict_file)
if len(all_predict_files) != 1:
raise ValueError('`eval_squad` only supports one predict file, '
'but got %s' % all_predict_files)
squad_model = get_squad_model_to_predict(strategy, bert_config,
init_checkpoint, input_meta_data)
all_predictions, all_nbest_json, scores_diff_json = prediction_output_squad(
strategy, input_meta_data, tokenizer, squad_lib, all_predict_files[0],
squad_model)
dump_to_files(all_predictions, all_nbest_json, scores_diff_json, squad_lib,
input_meta_data.get('version_2_with_negative', False))
with tf.io.gfile.GFile(FLAGS.predict_file, 'r') as reader:
dataset_json = json.load(reader)
pred_dataset = dataset_json['data']
if input_meta_data.get('version_2_with_negative', False):
eval_metrics = squad_evaluate_v2_0.evaluate(pred_dataset,
all_predictions,
scores_diff_json)
else:
eval_metrics = squad_evaluate_v1_1.evaluate(pred_dataset,
all_predictions)
return eval_metrics
def eval_squad(strategy,
input_meta_data,
tokenizer,
bert_config,
squad_lib,
init_checkpoint=None):
"""Get prediction results and evaluate them against ground truth."""
if init_checkpoint is None:
init_checkpoint = tf.train.latest_checkpoint(FLAGS.model_dir)
all_predictions, all_nbest_json, scores_diff_json = prediction_output_squad(
strategy, input_meta_data, tokenizer, bert_config, squad_lib,
init_checkpoint)
dump_to_files(all_predictions, all_nbest_json, scores_diff_json, squad_lib,
input_meta_data.get('version_2_with_negative', False))
with tf.io.gfile.GFile(FLAGS.predict_file, 'r') as reader:
dataset_json = json.load(reader)
pred_dataset = dataset_json['data']
if input_meta_data.get('version_2_with_negative', False):
eval_metrics = squad_evaluate_v2_0.evaluate(pred_dataset,
all_predictions,
scores_diff_json)
else:
eval_metrics = squad_evaluate_v1_1.evaluate(pred_dataset,
all_predictions)
return eval_metrics
def evaluate(self, model, input_fn, num_steps, eval_examples,
eval_features, predict_file, version_2_with_negative,
max_answer_length, null_score_diff_threshold, verbose_logging,
output_dir):
"""Evaluate QA model.
Args:
model: The model to be evaluated.
input_fn: Function that returns a tf.data.Dataset used for evaluation.
num_steps: Number of steps to evaluate the model.
eval_examples: List of `squad_lib.SquadExample` for evaluation data.
eval_features: List of `squad_lib.InputFeatures` for evaluation data.
predict_file: The input predict file.
version_2_with_negative: Whether the input predict file is SQuAD 2.0
format.
max_answer_length: The maximum length of an answer that can be generated.
This is needed because the start and end predictions are not conditioned
on one another.
null_score_diff_threshold: If null_score - best_non_null is greater than
the threshold, predict null. This is only used for SQuAD v2.
verbose_logging: If true, all of the warnings related to data processing
will be printed. A number of warnings are expected for a normal SQuAD
evaluation.
output_dir: The output directory to save output to json files:
predictions.json, nbest_predictions.json, null_odds.json. If None, skip
saving to json files.
Returns:
A dict contains two metrics: Exact match rate and F1 score.
"""
all_results = self.predict(model, input_fn, num_steps)
all_predictions, all_nbest_json, scores_diff_json = (
squad_lib.postprocess_output(
eval_examples,
eval_features,
all_results,
n_best_size=20,
max_answer_length=max_answer_length,
do_lower_case=self.do_lower_case,
version_2_with_negative=version_2_with_negative,
null_score_diff_threshold=null_score_diff_threshold,
verbose=verbose_logging))
if output_dir is not None:
dump_to_files(all_predictions, all_nbest_json, scores_diff_json,
version_2_with_negative, output_dir)
dataset_json = file_util.load_json_file(predict_file)
pred_dataset = dataset_json['data']
if version_2_with_negative:
eval_metrics = squad_evaluate_v2_0.evaluate(
pred_dataset, all_predictions, scores_diff_json)
else:
eval_metrics = squad_evaluate_v1_1.evaluate(
pred_dataset, all_predictions)
return eval_metrics
def reduce_aggregated_logs(self, aggregated_logs):
all_predictions, _, scores_diff = (self.squad_lib.postprocess_output(
self._eval_examples,
self._eval_features,
aggregated_logs,
self.task_config.n_best_size,
self.task_config.max_answer_length,
self.task_config.validation_data.do_lower_case,
version_2_with_negative=(
self.task_config.validation_data.version_2_with_negative),
null_score_diff_threshold=(
self.task_config.null_score_diff_threshold),
xlnet_format=self.task_config.validation_data.xlnet_format,
verbose=False))
with tf.io.gfile.GFile(self.task_config.validation_data.input_path,
'r') as reader:
dataset_json = json.load(reader)
pred_dataset = dataset_json['data']
if self.task_config.validation_data.version_2_with_negative:
eval_metrics = squad_evaluate_v2_0.evaluate(
pred_dataset, all_predictions, scores_diff)
# Filter out useless metrics, such as start_position_accuracy that
# we did not actually compute.
eval_metrics = {
'exact_match': eval_metrics['final_exact'],
'exact_match_threshold': eval_metrics['final_exact_thresh'],
'final_f1':
eval_metrics['final_f1'] / 100.0, # scale back to [0, 1].
'f1_threshold': eval_metrics['final_f1_thresh'],
'has_answer_exact_match': eval_metrics['HasAns_exact'],
'has_answer_f1': eval_metrics['HasAns_f1']
}
else:
eval_metrics = squad_evaluate_v1_1.evaluate(
pred_dataset, all_predictions)
# Filter out useless metrics, such as start_position_accuracy that
# we did not actually compute.
eval_metrics = {
'exact_match': eval_metrics['exact_match'],
'final_f1': eval_metrics['final_f1']
}
return eval_metrics