def main():
tf.reset_default_graph()
with tf.Session() as sess:
embeddings = generate_embedding_mat(vocab_size, embedding_size)
embed_file = './data/emb_mat.pkl'
if not os.path.exists(embed_file):
save_embedding_info(embed_file)
classifier = ModelAPI(model_dir, embed_path)
classifier.load_config()
classifier.config["token_emb_mat"] = embeddings
model = SelfAttRNN()
classifier.build_graph(sess, model, "/gpu:2")
actor = ActorNetwork(sess, lstm_dim, optimizer, lr, embeddings)
saver = tf.train.Saver()
model_file = "./checkpoints/{}".format(model_name)
restore_model(sess, saver, model_file)
# params = get_model_params(sess)
# get_simplify(sess, actor)
epoch = 5
try:
for e in range(epoch):
if use_RL:
train(sess,
actor,
classifier,
batchsize,
classifier_trainable=True)
else:
train_classifier(sess, classifier)
saver.save(sess, model_file)
except KeyboardInterrupt:
print(
'[INFO] Interrupt manually, try saving checkpoint for now...')
saver.save(sess, model_file)
def build_placeholder(self, config):
self.config = config
self.token_emb_mat = self.config["token_emb_mat"]
self.vocab_size = int(self.config["vocab_size"])
self.max_length = int(self.config["max_length"])
self.emb_size = int(self.config["emb_size"])
self.extra_symbol = self.config["extra_symbol"]
self.scope = self.config["scope"]
self.num_features = int(self.config["num_features"])
self.num_classes = int(self.config["num_classes"])
self.ema = self.config.get("ema", False)
self.grad_clipper = float(self.config.get("grad_clipper", 10.0))
print("--------vocab size---------", self.vocab_size)
print("--------max length---------", self.max_length)
print("--------emb size-----------", self.emb_size)
print("--------extra symbol-------", self.extra_symbol)
print("--------emb matrix---------", self.token_emb_mat.shape)
# ---- place holder -----
self.sent1_token = tf.placeholder(tf.int32, [None, None],
name='sent1_token')
self.gold_label = tf.placeholder(tf.int32, [None], name='gold_label')
self.sent1_len = tf.placeholder(tf.int32, [None],
name='sent1_token_length')
self.is_train = tf.placeholder(tf.bool, [], name='is_train')
self.features = tf.placeholder(tf.float32,
shape=[None, self.num_features],
name="features")
self.one_hot_label = tf.one_hot(self.gold_label, 2)
# ------------ other ---------
self.sent1_token_mask = tf.cast(self.sent1_token, tf.bool)
self.sent1_token_len = tf.reduce_sum(
tf.cast(self.sent1_token_mask, tf.int32), -1)
# ---------------- for dynamic learning rate -------------------
self.learning_rate = tf.placeholder(tf.float32, [], 'learning_rate')
self.learning_rate_value = float(self.config["learning_rate"])
self.dropout_keep_prob = tf.placeholder(tf.float32,
name="dropout_keep_prob")
self.emb_mat = generate_embedding_mat(self.vocab_size,
emb_len=self.emb_size,
init_mat=self.token_emb_mat,
extra_symbol=self.extra_symbol,
scope='gene_token_emb_mat')
self.s1_emb = tf.nn.embedding_lookup(self.emb_mat,
self.sent1_token) # bs,sl1,tel
self.global_step = tf.Variable(0, name="global_step", trainable=False)
self.learning_rate_updated = False
# ------ start ------
self.pred_probs = None
self.logits = None
self.loss = None
self.accuracy = None
self.var_ema = None
self.ema = None
self.opt = None
self.train_op = None