def __init__(self, args):
super(model_class, self).__init__()
self.args = args
self.gpu = args.cuda
#导入各种数据
input_data = data_generator.input_data(args=self.args)
#input_data.gen_het_rand_walk()
self.input_data = input_data
if self.args.train_test_label == 2: #为每个node生成邻居
#重启策略的随机游走，为每个节点采样固定数量的强相关的异构邻居，然后按类型分组
#任意节点开始随机游走，以p概率返回。采样到固定数量后就停止。
#为了采样采样的邻居包含所有类型的节点，不同类型节点的数量是受限的。
#对每个类型都选出按频率的topk邻居
input_data.het_walk_restart()
print("neighbor set generation finish")
exit(0)
#p是论文，a是作者，v是地点，然后可以组成一堆特征
feature_list = [input_data.p_abstract_embed, input_data.p_title_embed,\
input_data.p_v_net_embed, input_data.p_a_net_embed, input_data.p_ref_net_embed,\
input_data.p_net_embed, input_data.a_net_embed, input_data.a_text_embed,\
input_data.v_net_embed, input_data.v_text_embed]
for i in range(len(feature_list)):
feature_list[i] = torch.from_numpy(np.array(
feature_list[i])).float()
if self.gpu:
for i in range(len(feature_list)):
feature_list[i] = feature_list[i].cuda()
#self.feature_list = feature_list
#各自的邻居列表
a_neigh_list_train = input_data.a_neigh_list_train
p_neigh_list_train = input_data.p_neigh_list_train
v_neigh_list_train = input_data.v_neigh_list_train
a_train_id_list = input_data.a_train_id_list
p_train_id_list = input_data.p_train_id_list
v_train_id_list = input_data.v_train_id_list
self.model = tools.HetAgg(args, feature_list, a_neigh_list_train, p_neigh_list_train, v_neigh_list_train,\
a_train_id_list, p_train_id_list, v_train_id_list)#实例化model，tools会对异构的信息进行聚合
if self.gpu:
self.model.cuda()
self.parameters = filter(lambda p: p.requires_grad,
self.model.parameters())
self.optim = optim.Adam(self.parameters,
lr=self.args.lr,
weight_decay=0) #Adam优化器
self.model.init_weights()
def __init__(self, args):
super(model_class, self).__init__()
self.args = args
self.gpu = args.cuda
input_data = data_generator.input_data(args = self.args)
#input_data.gen_het_rand_walk()
self.input_data = input_data
if self.args.train_test_label == 2: #generate neighbor set of each node
input_data.het_walk_restart()
print ("neighbor set generation finish")
exit(0)
if self.args.train_test_label == 3: #generate neighbor according to metapath and rwr
input_data.mix_walk_restart()
print ("neighbor set generation finish")
exit(0)
if self.args.train_test_label == 4: #generate neighbor according to metapath
input_data.meta_walk()
print ("neighbor set generation finish")
exit(0)
feature_list = [input_data.p_abstract_embed, input_data.p_title_embed,\
input_data.p_v_net_embed, input_data.p_a_net_embed, input_data.p_ref_net_embed,\
input_data.p_net_embed, input_data.a_net_embed, input_data.a_text_embed,\
input_data.v_net_embed, input_data.v_text_embed]
for i in range(len(feature_list)):
feature_list[i] = torch.from_numpy(np.array(feature_list[i])).float()
if self.gpu:
for i in range(len(feature_list)):
feature_list[i] = feature_list[i].cuda()
#self.feature_list = feature_list
a_neigh_list_train = input_data.a_neigh_list_train
p_neigh_list_train = input_data.p_neigh_list_train
v_neigh_list_train = input_data.v_neigh_list_train
a_train_id_list = input_data.a_train_id_list
p_train_id_list = input_data.p_train_id_list
v_train_id_list = input_data.v_train_id_list
self.model = tools.HetAgg(args, feature_list, a_neigh_list_train, p_neigh_list_train, v_neigh_list_train,\
a_train_id_list, p_train_id_list, v_train_id_list)
if self.gpu:
self.model.cuda()
self.parameters = filter(lambda p: p.requires_grad, self.model.parameters())
self.optim = optim.Adam(self.parameters, lr=self.args.lr, weight_decay = 0)
self.model.init_weights()
c_reg = args.c_reg
margin_d = args.margin_d
c_tradeoff = args.c_tradeoff
batch_s = args.batch_size
lr = args.learn_rate
iter_max = args.train_iter_max
save_freq = args.save_model_freq
data_path = args.data_path
model_path = args.model_path
train_test_label = args.train_test_label
# input data and pre-train word embeddings
input_data = data_generator.input_data(args=args)
word_embed = input_data.word_embed
# generate negative paper ids in evaluation
if train_test_label == 2:
input_data.gen_evaluate_neg_ids()
# TSR model (objective function formulation) begin #
if train_test_label == 0:
# tensor preparation
# direct and indirect relation triples
a_p_p_dir = tf.placeholder(tf.int32, [None, 3])
a_p_p_indir = tf.placeholder(tf.int32, [None, 3])
# paper content in direct and indirect relations
p_c_dir_input = tf.placeholder(tf.int32, [None, c_len])
p_c_indir_input = tf.placeholder(tf.int32, [None, c_len])