def _forward(self, is_training, split_placeholders, **kwargs):
encoder = ALBERTEncoder(
config=self.albert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
token_type_ids=split_placeholders['segment_ids'],
scope='bert',
drop_pooler=self._drop_pooler,
**kwargs)
decoder = ALBERTDecoder(
albert_config=self.albert_config,
is_training=is_training,
encoder=encoder,
masked_lm_positions=split_placeholders['masked_lm_positions'],
masked_lm_ids=split_placeholders['masked_lm_ids'],
masked_lm_weights=split_placeholders['masked_lm_weights'],
sentence_order_labels=\
split_placeholders.get('sentence_order_labels'),
sample_weight=split_placeholders.get('sample_weight'),
scope_lm='cls/predictions',
scope_cls='cls/seq_relationship',
name='SOP',
**kwargs)
(total_loss, losses, probs, preds) = decoder.get_forward_outputs()
return (total_loss, losses, probs, preds)
def _get_encoder(model_name):
if model_name == 'bert' or model_name == 'roberta':
sketchy_encoder = BERTEncoder(
bert_config=self.bert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
segment_ids=split_placeholders['segment_ids'],
scope='bert',
**kwargs)
elif model_name == 'albert':
sketchy_encoder = ALBERTEncoder(
albert_config=self.bert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
segment_ids=split_placeholders['segment_ids'],
scope='bert',
**kwargs)
elif model_name == 'electra':
sketchy_encoder = BERTEncoder(
bert_config=self.bert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
segment_ids=split_placeholders['segment_ids'],
scope='electra',
**kwargs)
return sketchy_encoder
def _forward(self, is_training, split_placeholders, **kwargs):
encoder = ALBERTEncoder(albert_config=self.albert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
segment_ids=split_placeholders['segment_ids'],
scope='bert',
**kwargs)
encoder_output = encoder.get_sequence_output()
decoder = MRCDecoder(
is_training=is_training,
input_tensor=encoder_output,
label_ids=split_placeholders['label_ids'],
sample_weight=split_placeholders.get('sample_weight'),
scope='mrc',
**kwargs)
(total_loss, losses, probs, preds) = decoder.get_forward_outputs()
return (total_loss, losses, probs, preds)
def _forward(self, is_training, split_placeholders, **kwargs):
encoder = ALBERTEncoder(albert_config=self.albert_config,
is_training=is_training,
input_ids=split_placeholders['input_ids'],
input_mask=split_placeholders['input_mask'],
segment_ids=split_placeholders['segment_ids'],
scope='bert',
drop_pooler=self._drop_pooler,
**kwargs)
encoder_output = encoder.get_pooled_output()
decoder = CLSDecoder(
is_training=is_training,
input_tensor=encoder_output,
label_ids=split_placeholders['label_ids'],
label_size=self.label_size,
sample_weight=split_placeholders.get('sample_weight'),
scope='cls/seq_relationship',
**kwargs)
(total_loss, losses, probs, preds) = decoder.get_forward_outputs()
return (total_loss, losses, probs, preds)