def __init__(self, keep_tokens, load_path=None):
keras.backend.clear_session()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
keras.backend.set_session(tf.Session(config=gpu_config))
self.session = keras.backend.get_session()
need_load = False
if load_path and os.path.exists(load_path):
need_load = True
bert = build_transformer_model(
join(BERT_PATH, 'bert_config.json'),
None if need_load else join(BERT_PATH, 'bert_model.ckpt'),
keep_tokens=keep_tokens, # 只保留keep_tokens中的字，精简原字表
return_keras_model=False,
)
layers_out_lambda = Lambda(lambda x: x[:, 0])
layers_out_dense = Dense(units=2,
activation='softmax',
kernel_initializer=bert.initializer)
output = layers_out_lambda(bert.model.output)
output = layers_out_dense(output)
self.model = keras.models.Model(bert.model.input,
output,
name='Final-Model')
if need_load:
logger.info('=' * 15 +
'Load from checkpoint: {}'.format(load_path))
self.model.load_weights(load_path)
self.data_deal = None
def __init__(self, keep_tokens, num_classes, load_path=None):
keras.backend.clear_session()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
keras.backend.set_session(tf.Session(config=gpu_config))
self.session = keras.backend.get_session()
need_load = False
if load_path and os.path.exists(load_path):
need_load = True
bert = build_transformer_model(
config_path=join(BERT_PATH, 'bert_config.json'),
checkpoint_path=None if need_load else join(BERT_PATH, 'bert_model.ckpt'),
return_keras_model=False,
keep_tokens=keep_tokens,
)
output = Lambda(lambda x: x[:, 0])(bert.model.output)
output = Dense(
units=num_classes,
activation='softmax',
kernel_initializer=bert.initializer
)(output)
self.model = keras.models.Model(bert.model.input, output)
# self.model.summary()
if need_load:
logger.info('=' * 15 + 'Load from checkpoint: {}'.format(load_path))
self.model.load_weights(load_path)
self.data_deal = None
def __init__(self):
keras.backend.clear_session()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
keras.backend.set_session(tf.Session(config=gpu_config))
self.session = keras.backend.get_session()
self.tokenizer = Tokenizer(join(BERT_PATH, 'vocab.txt'),
do_lower_case=True) # 建立分词器
self.model = build_transformer_model(
join(BERT_PATH, 'bert_config.json'),
join(BERT_PATH, 'bert_model.ckpt'),
)
def __init__(self, keep_tokens, load_path=None):
keras.backend.clear_session()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
keras.backend.set_session(tf.Session(config=gpu_config))
self.session = keras.backend.get_session()
need_load = False
if load_path and os.path.exists(load_path):
need_load = True
self.model = build_transformer_model(
join(BERT_PATH, 'bert_config.json'),
None if need_load else join(BERT_PATH, 'bert_model.ckpt'),
application='lm',
keep_tokens=keep_tokens, # 只保留keep_tokens中的字，精简原字表
)
self.model.summary()
if need_load:
logger.info('=' * 15 +
'Load from checkpoint: {}'.format(load_path))
self.model.load_weights(load_path)
def __init__(self, tag='d', is_predict=False, load_path=None):
self.save_path = join(MODEL_PATH, 'rc_' + tag, 'trained.h5')
keras.backend.clear_session()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
self.session = tf.Session(config=gpu_config)
keras.backend.set_session(self.session)
self.data_deal = RCInput()
self.config_path = join(BERT_PATH, 'bert_config.json')
self.checkpoint_path = join(BERT_PATH, 'bert_model.ckpt')
self.tokenizer = self.data_deal.tokenizer
self.max_p_len = self.data_deal.max_p_len
self.max_q_len = self.data_deal.max_q_len
self.max_a_len = self.data_deal.max_a_len
self.batch_size = self.data_deal.batch_size
model = build_transformer_model(
self.config_path,
None if is_predict else self.checkpoint_path,
model='bert',
with_mlm=True,
keep_tokens=self.data_deal.keep_tokens, # 只保留keep_tokens中的字，精简原字表
)
output = Lambda(lambda x: x[:, 1:self.max_a_len + 1])(model.output)
self.model = Model(model.input, output)
# self.model.summary()
if load_path:
logger.info('Load from init checkpoint {} .'.format(load_path))
self.model.load_weights(load_path)
elif os.path.exists(self.save_path):
logger.info('Load from init checkpoint {} .'.format(
self.save_path))
self.model.load_weights(self.save_path)