def run_app_predictor(config):
try:
if config.model_name == "feat_ext_bert":
predictor = AppPredictor(config,
input_keys=["input_ids", "input_mask", "segment_ids"],
output_keys=["pool_output", "first_token_output", "all_hidden_outputs"],
job_name="ez_bert_feat")
predictor.run_predict()
elif config.model_name in ["text_comprehension_bert", "text_comprehension_bert_hae"]:
input_keys = ["input_ids", "input_mask", "segment_ids"] if config.model_name == "text_comprehension_bert" \
else ["input_ids", "input_mask", "segment_ids", "history_answer_marker"]
predictor = AppPredictor(config,
input_keys=input_keys,
output_keys=["start_logits", "end_logits"],
job_name=config.model_name + "_predictor")
preprocessor = preprocessors.get_preprocessor(
config.pretrain_model_name_or_path,
thread_num=predictor.thread_num,
input_queue=queue.Queue(),
output_queue=queue.Queue(),
preprocess_batch_size=config.predict_batch_size,
user_defined_config=config,
app_model_name=config.model_name)
postprocessor = postprocessors.get_postprocessors(
n_best_size=int(config.n_best_size) if hasattr(config, "n_best_size") else 20,
max_answer_length=int(config.max_answer_length) if hasattr(config, "max_answer_length") else 30,
output_schema=config.output_schema,
app_model_name=config.model_name,
thread_num=predictor.thread_num,
input_queue=queue.Queue(),
output_queue=queue.Queue())
predictor.run_predict(preprocessor=preprocessor, postprocessor=postprocessor)
elif config.model_name in ["text_match_dam", "text_match_damplus", "text_match_bicnn",
"text_match_hcnn", "text_classify_cnn"]:
predictor = AppPredictor(config,
input_keys=["input_ids_a", "input_mask_a", "input_ids_b", "input_mask_b"],
output_keys=["predictions", "probabilities", "logits"],
job_name=config.model_name + "_predictor")
preprocessor = DeepTextPreprocessor(config,
thread_num=predictor.thread_num,
input_queue=queue.Queue(),
output_queue=queue.Queue(),
job_name=config.model_name + "_predictor")
predictor.run_predict(preprocessor=preprocessor)
elif config.model_name in ["text_match_bert_two_tower"]:
raise NotImplementedError
else:
predictor = AppPredictor(config,
input_keys=["input_ids", "input_mask", "segment_ids"],
output_keys=["predictions", "probabilities", "logits"],
job_name=config.model_name + "_predictor")
predictor.run_predict()
except Exception as e:
traceback.print_exc(file=sys.stderr)
raise RuntimeError(str(e))
def build_logits(self, features, mode=None):
""" Building DAM text match graph
Args:
features (`OrderedDict`): A dict mapping raw input to tensors
mode (`bool`): tell the model whether it is under training
Returns:
logits (`Tensor`): The output after the last dense layer. Shape of [None, num_labels]
label_ids (`Tensor`): label_ids, shape of [None]
"""
text_preprocessor = DeepTextPreprocessor(self.config, mode=mode)
text_a_indices, text_a_masks, text_b_indices, text_b_masks, label_ids = text_preprocessor(
features)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
word_embeddings = self._add_word_embeddings(
vocab_size=text_preprocessor.vocab.size,
embed_size=self.config.embedding_size,
pretrained_word_embeddings=text_preprocessor.
pretrained_word_embeddings,
trainable=not self.config.fix_embedding)
a_embeds = tf.nn.embedding_lookup(word_embeddings, text_a_indices)
b_embeds = tf.nn.embedding_lookup(word_embeddings, text_b_indices)
dam_output_features = layers.DAMEncoder(self.config.hidden_size)(
[a_embeds, b_embeds, text_a_masks, text_b_masks],
training=is_training)
dam_output_features = tf.layers.dropout(
dam_output_features,
rate=0.2,
training=is_training,
name='dam_out_features_dropout')
dam_output_features = layers.Dense(
self.config.hidden_size,
activation=tf.nn.relu,
kernel_initializer=layers.get_initializer(0.02),
name='dam_out_features_projection')(dam_output_features)
logits = layers.Dense(self.config.num_labels,
kernel_initializer=layers.get_initializer(0.02),
name='output_layer')(dam_output_features)
self.check_and_init_from_checkpoint(mode)
return logits, label_ids
def build_logits(self, features, mode=None):
""" Building DAM text match graph
Args:
features (`OrderedDict`): A dict mapping raw input to tensors
mode (`bool`): tell the model whether it is under training
Returns:
logits (`Tensor`): The output after the last dense layer. Shape of [None, num_labels]
label_ids (`Tensor`): label_ids, shape of [None]
"""
text_preprocessor = DeepTextPreprocessor(self.config, mode=mode)
text_indices, text_masks, _, _, label_ids = text_preprocessor(features)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
word_embeddings = self._add_word_embeddings(
vocab_size=text_preprocessor.vocab.size,
embed_size=self.config.embedding_size,
pretrained_word_embeddings=text_preprocessor.
pretrained_word_embeddings,
trainable=not self.config.fix_embedding)
text_embeds = tf.nn.embedding_lookup(word_embeddings, text_indices)
output_features = layers.TextCNNEncoder(
num_filters=self.config.num_filters,
filter_sizes=self.config.filter_sizes,
embed_size=self.config.embedding_size,
max_seq_len=self.config.sequence_length,
)([text_embeds, text_masks], training=is_training)
output_features = tf.layers.dropout(output_features,
rate=self.config.dropout_rate,
training=is_training,
name='output_features')
logits = layers.Dense(self.config.num_labels,
kernel_initializer=layers.get_initializer(0.02),
name='output_layer')(output_features)
self.check_and_init_from_checkpoint(mode)
return logits, label_ids
def build_logits(self, features, mode=None):
text_preprocessor = DeepTextPreprocessor(self.config, mode=mode)
text_a_indices, text_a_masks, text_b_indices, text_b_masks, label_ids = text_preprocessor(
features)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
word_embeddings = self._add_word_embeddings(
vocab_size=text_preprocessor.vocab.size,
embed_size=self.config.embedding_size,
pretrained_word_embeddings=text_preprocessor.
pretrained_word_embeddings,
trainable=not self.config.fix_embedding)
a_embeds = tf.nn.embedding_lookup(word_embeddings, text_a_indices)
b_embeds = tf.nn.embedding_lookup(word_embeddings, text_b_indices)
hcnn_output_features = layers.HybridCNNEncoder(
num_filters=self.config.hidden_size,
l2_reg=self.config.l2_reg,
filter_size=self.config.filter_size)(
[a_embeds, b_embeds, text_a_masks, text_b_masks])
hcnn_output_features = tf.layers.dropout(
hcnn_output_features,
rate=0.2,
training=is_training,
name='dam_out_features_dropout')
hcnn_output_features = layers.Dense(
self.config.hidden_size,
activation=tf.nn.relu,
kernel_initializer=layers.get_initializer(0.02),
name='dam_out_features_projection')(hcnn_output_features)
logits = layers.Dense(self.config.num_labels,
kernel_initializer=layers.get_initializer(0.02),
name='output_layer')(hcnn_output_features)
self.check_and_init_from_checkpoint(mode)
return logits, label_ids