cnn_setting["num_classes"] = len(properties_list)
cnn_setting["dim_pos"] = int(c_para.get("Global", "pos_dim"))
cnn_setting["num_filters"] = int(c_para.get("CNN", "num_filters"))
policy_setting = {}
policy_setting["state_dim"] = cnn_setting[
"num_filters"] * 2 + dim_entity * 2 + dim_relation
policy_setting["learning_rate"] = float(c_para.get("RL", "learning_rate"))
assign_rate = float(c_para.get("RL", "assign_rate"))
log_file = c_path.get("RL", "log_file")
log_fo = open(log_file, 'w', encoding='utf-8')
del (entity2id_dict, entities_list)
best_reward = -100000
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
cnn_main = cnn_class.CNN(cnn_setting, cnn_main_net)
cnn_target = cnn_class.CNN(cnn_setting, cnn_target_net)
cnn_best = cnn_class.CNN(cnn_setting, cnn_best_net_name)
cnn_main.build_model()
cnn_target.build_model()
cnn_best.build_model()
agent_main = rl_class.Agent(policy_setting, policy_main_net_name)
agent_target = rl_class.Agent(policy_setting, policy_target_net_name)
agent_best = rl_class.Agent(policy_setting, policy_best_net_name)
agent_main.build_model()
agent_target.build_model()
agent_best.build_model()
sess.run(tf.global_variables_initializer())
utils.load_model(sess, pre_cnn_path, cnn_main_net)
utils.load_model(sess, pre_rl_path, policy_main_net_name)
sess.run(cnn_main.WV.assign(wv_npy))
cnn_setting["vocab_size"] = vocab_size
policy_setting = {}
policy_setting["state_dim"] = cnn_setting[
"num_filters"] * 2 + dim_entity * 2 + dim_relation
policy_setting["learning_rate"] = learning_rate
properties_list = fo.load_relations(relation_data)
positionVec_npy = fo.load_pos_vec(position_vec_path)
entity2id_dict, entity_vec_npy = fo.load_entity2vec(entity_id_path,
entity_vec_path)
sents_list, en1_position_list, en2_position_list, y_list, Bags = fo.load_agent_test_data(
test_data, cnn_setting["max_sequence_length"], num_classes, vocab_dict,
properties_list, entity2id_dict)
selected_sents = []
with tf.Session() as sess:
cnn = cnn_class.CNN(cnn_setting, "cnn_target")
agent = rl_class.Agent(policy_setting, "policy_target")
cnn.build_model()
agent.build_model()
sess.run(tf.global_variables_initializer())
utils.load_model(sess, cnn_model_path, "cnn_target")
utils.load_model(sess, agent_model_path, "policy_target")
sess.run(cnn.WV.assign(wv_npy))
sess.run(cnn.POS.assign(positionVec_npy))
env = env_class.Environment(sess,
cnn,
sents_list,
y_list,
en1_position_list,
en2_position_list,
entity_vec_npy,
# print(len(vocab_dict.keys()),len(wv_npy))
# vocab_size = len(vocab_dict.keys())
# sents_list,en1_position_list,en2_position_list,lbox_infos,dict_data,e1_list,e2_list = fo.load_extract_data(input_data,cnn_setting["max_sequence_length"],vocab_dict)
vocab_dict, wv_npy, sents_list, en1_position_list, en2_position_list, e1_list, e2_list, lbox_ids, dict_data = fo.load_extract_data(
input_data, w2v_data, cnn_setting["dim_word"],
cnn_setting["max_sequence_length"])
vocab_size = len(vocab_dict.keys())
num_classes = len(properties_list)
positionVec_npy = fo.load_pos_vec(position_vec_path)
total_data_size = len(sents_list)
print("[DONE] Loading data")
print("[BUILD] CNN model")
cnn_setting["vocab_size"] = vocab_size
with tf.Session() as sess:
cnn = cnn_class.CNN(cnn_setting, net_name)
cnn.build_model()
sess.run(tf.global_variables_initializer())
utils.load_model(sess, model_path, net_name)
sess.run(cnn.WV.assign(wv_npy))
sess.run(cnn.POS.assign(positionVec_npy))
total_batch_size = int(total_data_size / batch_size) + 1
predicted_y = []
probs = []
for batch in tqdm(range(total_batch_size)):
st = batch * batch_size
en = min((batch + 1) * batch_size, total_data_size)
batch_sents = sents_list[st:en]
batch_pos1 = en1_position_list[st:en]
batch_pos2 = en2_position_list[st:en]