def model_fn(features, labels, mode, params):
logger.info("*** Features ***")
if isinstance(features, dict):
features = features['words'], features['words_seq'], features[
'text_length']
input_ids, input_word_ids, text_length_list = features
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tag_model = BILSTMONLY(params)
# def metric_fn(label_ids, pred_ids):
# return {
# 'precision': precision(label_ids, pred_ids, params["num_labels"]),
# 'recall': recall(label_ids, pred_ids, params["num_labels"]),
# 'f1': f1(label_ids, pred_ids, params["num_labels"])
# }
#
# eval_metrics = metric_fn(labels, pred_ids)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
loss, pred_ids, weight = tag_model(input_ids, input_word_ids,
labels, text_length_list,
is_training)
train_op = optimization.create_optimizer(loss, args.lr,
params["decay_steps"],
args.clip_norm)
hook_dict = {}
# precision_score, precision_update_op = precision(labels=labels, predictions=pred_ids,
# num_classes=params["num_labels"], weights=weight)
#
# recall_score, recall_update_op = recall(labels=labels,
# predictions=pred_ids, num_classes=params["num_labels"],
# weights=weight)
hook_dict['loss'] = loss
hook_dict['global_steps'] = tf.train.get_or_create_global_step()
logging_hook = tf.train.LoggingTensorHook(
hook_dict, every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[logging_hook])
elif mode == tf.estimator.ModeKeys.EVAL:
loss, pred_ids, weight = tag_model(input_ids, input_word_ids,
labels, text_length_list,
is_training)
# pred_ids = tf.argmax(logits, axis=-1, output_type=tf.int32)
# weight = tf.sequence_mask(text_length_list)
# precision_score, precision_update_op = precision(labels=labels,predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
#
# recall_score, recall_update_op =recall(labels=labels,
# predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
f1_score_val, f1_update_op_val = f1(
labels=labels,
predictions=pred_ids,
num_classes=params["num_labels"],
weights=weight,
average="micro")
# acc_score_val,acc_score_op_val = tf.metrics.accuracy(labels=labels,predictions=pred_ids,weights=weight)
eval_metric_ops = {"f1": (f1_score_val, f1_update_op_val)}
eval_hook_dict = {"f1": f1_score_val, "loss": loss}
eval_logging_hook = tf.train.LoggingTensorHook(
eval_hook_dict, at_end=True, every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
eval_metric_ops=eval_metric_ops,
mode=mode,
loss=loss,
evaluation_hooks=[eval_logging_hook])
else:
pred_ids = tag_model(input_ids, input_word_ids, labels,
text_length_list, is_training, True)
output_spec = tf.estimator.EstimatorSpec(mode=mode,
predictions=pred_ids)
return output_spec
def model_fn(features, labels, mode, params):
logger.info("*** Features ***")
if isinstance(features, dict):
features = features['words'], features['text_length'], features[
'query_length'], features['token_type_ids']
print(features)
input_ids, text_length_list, query_length_list, token_type_id_list = features
if labels is not None:
start_labels, end_labels = labels
else:
start_labels, end_labels = None, None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
is_testing = (mode == tf.estimator.ModeKeys.PREDICT)
bert_config = modeling.BertConfig.from_json_file(bert_config_file)
tag_model = bertMRC(params, bert_config)
# input_ids,labels,token_type_ids_list,query_len_list,text_length_list,is_training,is_testing=False
if is_testing:
pred_start_ids, pred_end_ids, weight = tag_model(
input_ids, start_labels, end_labels, token_type_id_list,
query_length_list, text_length_list, is_training, is_testing)
else:
loss, pred_start_ids, pred_end_ids, weight = tag_model(
input_ids, start_labels, end_labels, token_type_id_list,
query_length_list, text_length_list, is_training)
# def metric_fn(label_ids, pred_ids):
# return {
# 'precision': precision(label_ids, pred_ids, params["num_labels"]),
# 'recall': recall(label_ids, pred_ids, params["num_labels"]),
# 'f1': f1(label_ids, pred_ids, params["num_labels"])
# }
#
# eval_metrics = metric_fn(labels, pred_ids)
tvars = tf.trainable_variables()
# 加载BERT模型
if init_checkpoints:
(assignment_map, initialized_variable_names) = \
modeling.get_assignment_map_from_checkpoint(tvars,
init_checkpoints)
tf.train.init_from_checkpoint(init_checkpoints, assignment_map)
output_spec = None
# f1_score_val, f1_update_op_val = f1(labels=labels, predictions=pred_ids, num_classes=params["num_labels"],
# weights=weight)
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer(loss,
args.lr,
params["decay_steps"],
args.clip_norm,
use_tpu=False)
hook_dict = {}
# precision_score, precision_update_op = precision(labels=labels, predictions=pred_ids,
# num_classes=params["num_labels"], weights=weight)
#
# recall_score, recall_update_op = recall(labels=labels,
# predictions=pred_ids, num_classes=params["num_labels"],
# weights=weight)
hook_dict['loss'] = loss
hook_dict['global_steps'] = tf.train.get_or_create_global_step()
logging_hook = tf.train.LoggingTensorHook(
hook_dict, every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[logging_hook])
elif mode == tf.estimator.ModeKeys.EVAL:
# pred_ids = tf.argmax(logits, axis=-1, output_type=tf.int32)
# weight = tf.sequence_mask(text_length_list)
# precision_score, precision_update_op = precision(labels=labels,predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
#
# recall_score, recall_update_op =recall(labels=labels,
# predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
f1_start_val, f1_update_op_val = f1(labels=start_labels,
predictions=pred_start_ids,
num_classes=2,
weights=weight,
average="macro")
f1_end_val, f1_end_update_op_val = f1(labels=end_labels,
predictions=pred_end_ids,
num_classes=2,
weights=weight,
average="macro")
# f1_score_val_micro,f1_update_op_val_micro = f1(labels=labels,predictions=pred_ids,num_classes=params["num_labels"],weights=weight,average="micro")
# acc_score_val,acc_score_op_val = tf.metrics.accuracy(labels=labels,predictions=pred_ids,weights=weight)
# eval_loss = tf.metrics.mean_squared_error(labels=labels, predictions=pred_ids,weights=weight)
eval_metric_ops = {
"f1_start_micro": (f1_start_val, f1_update_op_val),
"f1_end_micro": (f1_end_val, f1_end_update_op_val)
}
# eval_hook_dict = {"f1":f1_score_val,"loss":loss}
# eval_logging_hook = tf.train.LoggingTensorHook(
# at_end=True,every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
eval_metric_ops=eval_metric_ops, mode=mode, loss=loss)
else:
output_spec = tf.estimator.EstimatorSpec(mode=mode,
predictions={
"start_ids":
pred_start_ids,
"end_ids":
pred_end_ids
})
return output_spec
def model_fn(features, labels, mode, params):
logger.info("*** Features ***")
if isinstance(features, dict):
features = features['words'], features['token_type_ids'], features[
'text_length']
# print(features)
# input_ids,token_type_ids,text_length_list = features
input_ids, token_type_ids, text_length_list = features
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
is_testing = (mode == tf.estimator.ModeKeys.PREDICT)
bert_config = modeling.BertConfig.from_json_file(bert_config_file)
# tag_model = bertEventType(params,bert_config)
tag_model = bertEventType(params, bert_config)
if is_testing:
# pred_ids = tag_model(input_ids, labels, text_length_list, token_type_ids,is_training,is_testing)
pred_ids = tag_model(input_ids, labels, text_length_list,
token_type_ids, is_training, is_testing)
else:
per_example_loss, loss, pred_ids = tag_model(
input_ids, labels, text_length_list, token_type_ids,
is_training)
tvars = tf.trainable_variables()
# 加载BERT模型
if init_checkpoints:
(assignment_map, initialized_variable_names) = \
modeling.get_assignment_map_from_checkpoint(tvars,
init_checkpoints)
tf.train.init_from_checkpoint(init_checkpoints, assignment_map)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer(loss, args.lr,
params["decay_steps"],
None, False)
hook_dict = {}
hook_dict['loss'] = loss
hook_dict['global_steps'] = tf.train.get_or_create_global_step()
logging_hook = tf.train.LoggingTensorHook(
hook_dict, every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[logging_hook])
elif mode == tf.estimator.ModeKeys.EVAL:
pred_ids = tf.where(pred_ids > 0.5, tf.ones_like(pred_ids),
tf.zeros_like(pred_ids))
print(pred_ids)
print(labels)
f1_score_val_micro, f1_update_op_val_micro = f1(
labels=labels, predictions=pred_ids, num_classes=2)
eval_metrics = {
"f1_score_micro": (f1_score_val_micro, f1_update_op_val_micro)
}
eval_metrics['eval_loss'] = tf.metrics.mean(
values=per_example_loss)
output_spec = tf.estimator.EstimatorSpec(
eval_metric_ops=eval_metrics, mode=mode, loss=loss)
else:
output_spec = tf.estimator.EstimatorSpec(mode=mode,
predictions=pred_ids)
return output_spec
def model_fn(features, labels, mode, params):
logger.info("*** Features ***")
if isinstance(features, dict):
features = features['words'], features['text_length'], features[
'query_length'], features['token_type_ids']
print(features)
input_ids, text_length_list, query_length_list, token_type_id_list = features
if labels is not None:
start_labels, end_labels, has_answer_label = labels
else:
start_labels, end_labels, has_answer_label = None, None, None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
is_testing = (mode == tf.estimator.ModeKeys.PREDICT)
bert_config = modeling.BertConfig.from_json_file(bert_config_file)
tag_model = VerfiyMRC(params, bert_config)
# input_ids,labels,token_type_ids_list,query_len_list,text_length_list,is_training,is_testing=False
if is_testing:
pred_start_ids, pred_end_ids, weight, predict_start_prob, predict_end_prob, has_answer_prob = tag_model(
input_ids, start_labels, end_labels, token_type_id_list,
query_length_list, text_length_list, has_answer_label,
is_training, is_testing)
# predict_ids,weight,predict_prob = tag_model(input_ids,labels,token_type_id_list,query_length_list,text_length_list,is_training,is_testing)
else:
loss, pred_start_ids, pred_end_ids, weight, predict_start_prob, predict_end_prob, has_answer_prob = tag_model(
input_ids, start_labels, end_labels, token_type_id_list,
query_length_list, text_length_list, has_answer_label,
is_training)
# loss,predict_ids,weight,predict_prob = tag_model(input_ids,labels,token_type_id_list,query_length_list,text_length_list,is_training,is_testing)
# def metric_fn(label_ids, pred_ids):
# return {
# 'precision': precision(label_ids, pred_ids, params["num_labels"]),
# 'recall': recall(label_ids, pred_ids, params["num_labels"]),
# 'f1': f1(label_ids, pred_ids, params["num_labels"])
# }
#
# eval_metrics = metric_fn(labels, pred_ids)
tvars = tf.trainable_variables()
# 加载BERT模型
if init_checkpoints:
(assignment_map, initialized_variable_names) = \
modeling.get_assignment_map_from_checkpoint(tvars,
init_checkpoints)
tf.train.init_from_checkpoint(init_checkpoints, assignment_map)
output_spec = None
# f1_score_val, f1_update_op_val = f1(labels=labels, predictions=pred_ids, num_classes=params["num_labels"],
# weights=weight)
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer(
loss, args.lr, params["train_steps"],
params["num_warmup_steps"], args.clip_norm)
hook_dict = {}
# precision_score, precision_update_op = precision(labels=labels, predictions=pred_ids,
# num_classes=params["num_labels"], weights=weight)
#
# recall_score, recall_update_op = recall(labels=labels,
# predictions=pred_ids, num_classes=params["num_labels"],
# weights=weight)
hook_dict['loss'] = loss
hook_dict['global_steps'] = tf.train.get_or_create_global_step()
logging_hook = tf.train.LoggingTensorHook(
hook_dict, every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[logging_hook])
elif mode == tf.estimator.ModeKeys.EVAL:
has_answer_pred = tf.where(has_answer_prob > 0.5,
tf.ones_like(has_answer_prob),
tf.zeros_like(has_answer_prob))
# pred_ids = tf.argmax(logits, axis=-1, output_type=tf.int32)
# weight = tf.sequence_mask(text_length_list)
# precision_score, precision_update_op = precision(labels=labels,predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
#
# recall_score, recall_update_op =recall(labels=labels,
# predictions=pred_ids,num_classes=params["num_labels"],weights=weight)
# def metric_fn(per_example_loss, label_ids, probabilities):
# logits_split = tf.split(probabilities, params["num_labels"], axis=-1)
# label_ids_split = tf.split(label_ids, params["num_labels"], axis=-1)
# # metrics change to auc of every class
# eval_dict = {}
# for j, logits in enumerate(logits_split):
# label_id_ = tf.cast(label_ids_split[j], dtype=tf.int32)
# current_auc, update_op_auc = tf.metrics.auc(label_id_, logits)
# eval_dict[str(j)] = (current_auc, update_op_auc)
# eval_dict['eval_loss'] = tf.metrics.mean(values=per_example_loss)
# return eval_dict
# eval_metrics = metric_fn(per_example_loss, labels, pred_ids)
f1_start_val, f1_update_op_val = f1(labels=start_labels,
predictions=pred_start_ids,
num_classes=2,
weights=weight,
average="macro")
f1_end_val, f1_end_update_op_val = f1(labels=end_labels,
predictions=pred_end_ids,
num_classes=2,
weights=weight,
average="macro")
# f1_val,f1_update_op_val = f1(labels=labels,predictions=predict_ids,num_classes=3,weights=weight,average="macro")
has_answer_label = tf.cast(has_answer_label, tf.float32)
f1_has_val, f1_has_update_op_val = f1(labels=has_answer_label,
predictions=has_answer_pred,
num_classes=2)
# f1_score_val_micro,f1_update_op_val_micro = f1(labels=labels,predictions=pred_ids,num_classes=params["num_labels"],weights=weight,average="micro")
# acc_score_val,acc_score_op_val = tf.metrics.accuracy(labels=labels,predictions=pred_ids,weights=weight)
# eval_loss = tf.metrics.mean_squared_error(labels=labels, predictions=pred_ids,weights=weight)
eval_metric_ops = {
"f1_start_macro": (f1_start_val, f1_update_op_val),
"f1_end_macro": (f1_end_val, f1_end_update_op_val),
"f1_has_answer_macro": (f1_has_val, f1_has_update_op_val),
"eval_loss": tf.metrics.mean(values=loss)
}
# eval_metric_ops = {
# "f1_macro":(f1_val,f1_update_op_val),
# "eval_loss":tf.metrics.mean(values=loss)}
# eval_hook_dict = {"f1":f1_score_val,"loss":loss}
# eval_logging_hook = tf.train.LoggingTensorHook(
# at_end=True,every_n_iter=args.print_log_steps)
output_spec = tf.estimator.EstimatorSpec(
eval_metric_ops=eval_metric_ops, mode=mode, loss=loss)
else:
output_spec = tf.estimator.EstimatorSpec(mode=mode,
predictions={
"start_ids":
pred_start_ids,
"end_ids":
pred_end_ids,
"start_probs":
predict_start_prob,
"end_probs":
predict_end_prob,
"has_answer_probs":
has_answer_prob
})
# output_spec = tf.estimator.EstimatorSpec(
# mode=mode,
# predictions={"pred_ids":predict_ids,"pred_probs":predict_prob}
# )
return output_spec