def __init__(self, args, dep_tag_num, pos_tag_num):
super(Aspect_Bert_GAT, self).__init__()
self.args = args
# Bert
config = BertConfig.from_pretrained(args.bert_model_dir)
self.bert = BertModel.from_pretrained(
args.bert_model_dir, config=config, from_tf =False)
self.dropout_bert = nn.Dropout(config.hidden_dropout_prob)
self.dropout = nn.Dropout(args.dropout)
args.embedding_dim = config.hidden_size # 768
if args.highway:
self.highway_dep = Highway(args.num_layers, args.embedding_dim)
self.highway = Highway(args.num_layers, args.embedding_dim)
gcn_input_dim = args.embedding_dim
# GAT
self.gat_dep = [RelationAttention(in_dim=args.embedding_dim).to(args.device) for i in range(args.num_heads)]
self.dep_embed = nn.Embedding(dep_tag_num, args.embedding_dim)
last_hidden_size = args.embedding_dim * 2
layers = [
nn.Linear(last_hidden_size, args.final_hidden_size), nn.ReLU()]
for _ in range(args.num_mlps - 1):
layers += [nn.Linear(args.final_hidden_size,
args.final_hidden_size), nn.ReLU()]
self.fcs = nn.Sequential(*layers)
self.fc_final = nn.Linear(args.final_hidden_size, args.num_classes)
def __init__(self, args, dep_tag_num, pos_tag_num):
super(Aspect_Text_GAT_ours, self).__init__()
self.args = args
num_embeddings, embed_dim = args.glove_embedding.shape
self.embed = nn.Embedding(num_embeddings, embed_dim)
self.embed.weight = nn.Parameter(args.glove_embedding,
requires_grad=False)
self.dropout = nn.Dropout(args.dropout)
self.tanh = nn.Tanh()
if args.highway:
self.highway_dep = Highway(args.num_layers, args.embedding_dim)
self.highway = Highway(args.num_layers, args.embedding_dim)
self.bilstm = nn.LSTM(input_size=args.embedding_dim,
hidden_size=args.hidden_size,
bidirectional=True,
batch_first=True,
num_layers=args.num_layers)
gcn_input_dim = args.hidden_size * 2
# if args.gat:
self.gat_dep = [
RelationAttention(in_dim=args.embedding_dim).to(args.device)
for i in range(args.num_heads)
]
if args.gat_attention_type == 'linear':
self.gat = [
LinearAttention(in_dim=gcn_input_dim,
mem_dim=gcn_input_dim).to(args.device)
for i in range(args.num_heads)
] # we prefer to keep the dimension unchanged
elif args.gat_attention_type == 'dotprod':
self.gat = [
DotprodAttention().to(args.device)
for i in range(args.num_heads)
]
else:
# reshaped gcn
self.gat = nn.Linear(gcn_input_dim, gcn_input_dim)
self.dep_embed = nn.Embedding(dep_tag_num, args.embedding_dim)
last_hidden_size = args.hidden_size * 4
layers = [
nn.Linear(last_hidden_size, args.final_hidden_size),
nn.ReLU()
]
for _ in range(args.num_mlps - 1):
layers += [
nn.Linear(args.final_hidden_size, args.final_hidden_size),
nn.ReLU()
]
self.fcs = nn.Sequential(*layers)
self.fc_final = nn.Linear(args.final_hidden_size, args.num_classes)