def get_init_ops(self):
init_ops = []
if self.conf.embedding_init:
init_ops = [
tf.variables_initializer(
set(self.optimizer_params + self.global_params +
self.trainable_params) - set([self.embedding]))
]
w2v = np.load(self.conf.embedding_init)
init_ops.append(self.embedding.assign(w2v))
else:
init_ops = [
tf.variables_initializer(
set(self.optimizer_params + self.global_params +
self.trainable_params))
]
if self.task_id == 0:
vocab_file = filter(lambda x: re.match("vocab\d+\.all", x) != None,
os.listdir(self.conf.data_dir))[0]
f = codecs.open(os.path.join(self.conf.data_dir, vocab_file))
k = [line.strip() for line in f]
k = k[0:self.conf.output_vocab_size]
v = [i for i in range(len(k))]
op_in = self.in_table.insert(
constant_op.constant(k), constant_op.constant(v,
dtype=tf.int64))
op_out = self.out_table.insert(
constant_op.constant(v, dtype=tf.int64),
constant_op.constant(k))
init_ops.extend([op_in, op_out])
return init_ops
def get_init_ops(self, job_type, task_id):
init_ops = [
tf.variables_initializer(
set(self.optimizer_params + self.global_params +
self.trainable_params))
]
if self.conf.embedding_init:
w2v = np.load(self.conf.embedding_init)
init_ops.append(self.embedding.assign(w2v))
init_ops = list(set(init_ops) - set([self.embedding]))
if self.conf.topic_embedding_init:
t2v = np.load(self.conf.topic_embedding_init)
init_ops.append(self.topic_embedding.assign(t2v))
init_ops = list(set(init_ops) - set([self.topic_embedding]))
if not self.for_deploy and task_id == 0:
vocab_file = filter(
lambda x: re.match("vocab[0-9]+\.all", x) != None,
os.listdir(self.conf.data_dir))[0]
f = codecs.open(os.path.join(self.conf.data_dir, vocab_file))
k = [line.strip() for line in f]
k = k[0:self.conf.input_vocab_size]
v = [i for i in range(len(k))]
op_in = self.in_table.insert(
constant_op.constant(k), constant_op.constant(v,
dtype=tf.int64))
op_out = self.out_table.insert(
constant_op.constant(v, dtype=tf.int64),
constant_op.constant(k))
# Topic
topic_vocab_file = "vocab.topic"
ft = codecs.open(os.path.join(self.conf.data_dir,
topic_vocab_file))
k = [line.strip() for line in ft]
v = [i for i in range(len(k))]
op_topic_in = self.topic_in_table.insert(
constant_op.constant(k), constant_op.constant(v,
dtype=tf.int64))
init_ops.extend([op_in, op_out, op_topic_in])
return init_ops
