def __init__(self, config):
self.graph = tf.Graph()
self.config = config
self.A_Net = Attr_Net(self.config)
self.S_Net = Senti_Net(self.config)
'train_data_path': os.path.join(model_data_path, 'train_han_fasttext.pkl'),
'dev_data_path': os.path.join(model_data_path, 'dev_han_fasttext.pkl'),
'testa_data_path': os.path.join(model_data_path, 'testa_han_fasttext.pkl'),
"attr_init_model": '/hdd/lujunyu/model/meituan/D_HAN_MC/attr/',
"senti_init_model":
'/hdd/lujunyu/model/meituan/D_HAN_MC/senti/', #should be set for test
"rand_seed": 1,
"learning_rate": 1e-3,
"attribute_threshold": 0.5,
"vocab_size": 266078, #111695
"emb_dim": 300,
"batch_size": 1, #200 for test
"max_rev_len": 25,
"max_sent_len": 251,
'attribute_num': 20,
'sentiment_num': 3,
"max_to_keep": 1,
"num_scan_data": 5,
"_EOS_":
0, #1455 for DSTC7, 28270 for DAM_source, #1 for douban data , 6 for advising
"rnn_dim": 300,
'bert_dim': 768 * 2,
'multi_head': 8,
'Model': 'D_HAN_MC'
}
conf.update({'save_path': os.path.join(model_path, conf['Model'] + '/senti/')})
attr_model = Attr_Net(conf)
senti_model = Senti_Net(conf)
sent_test.test(conf, attr_model, senti_model)
class Joint_Net(object):
def __init__(self, config):
self.graph = tf.Graph()
self.config = config
self.A_Net = Attr_Net(self.config)
self.S_Net = Senti_Net(self.config)
def build_graph(self):
if self.config['rand_seed'] is not None:
rand_seed = self.config['rand_seed']
tf.set_random_seed(rand_seed)
print('set tf random seed: %s' % self.config['rand_seed'])
with self.graph.as_default():
self.review, self.attr_label, self.senti_label, self.is_training, self.table = layers.joint_net_input(
self.config)
# convert table to variable
table_v = tf.Variable(self.table, name='table')
# review_embed : (batch, rev_len, sent_len, emb_dim)
review_embed = tf.nn.embedding_lookup(table_v, self.review)
# rev_len:(batch,)
# sent_len:(batch, rev)
rev_len, sent_len = op.generate_mask(self.config, self.review)
with tf.variable_scope('attr_net', reuse=tf.AUTO_REUSE):
self.A_Net.build(review_embed, rev_len, sent_len,
self.attr_label)
with tf.variable_scope('senti_net', reuse=tf.AUTO_REUSE):
self.S_Net.build(review_embed, rev_len, sent_len,
self.A_Net.doc_att, self.attr_label,
self.senti_label)
self.joint_loss = self.A_Net.loss + self.S_Net.loss
# Calculate cross-entropy loss
tv = tf.trainable_variables()
for v in tv:
print(v)
self.global_step = tf.Variable(0, trainable=False)
initial_learning_rate = self.config['learning_rate']
self.learning_rate = tf.train.exponential_decay(
initial_learning_rate,
global_step=self.global_step,
decay_steps=300,
decay_rate=0.9,
staircase=True)
Optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.optimizer = Optimizer.minimize(self.joint_loss,
global_step=self.global_step)
self.init = tf.global_variables_initializer()
self.saver = tf.train.Saver(max_to_keep=self.config["max_to_keep"])
self.all_variables = tf.global_variables()
self.grads_and_vars = Optimizer.compute_gradients(self.joint_loss)
for grad, var in self.grads_and_vars:
if grad is None:
print(var)
self.capped_gvs = [(tf.clip_by_value(grad, -1, 1), var)
for grad, var in self.grads_and_vars]
self.g_updates = Optimizer.apply_gradients(
self.capped_gvs, global_step=self.global_step)
return self.graph