def __init__(self, config, vocab_size, embedding=None):
super(pointer_decoder, self).__init__()
if embedding is not None:
self.embedding = embedding
else:
self.embedding = nn.Embedding(vocab_size, config.emb_size)
self.rnn = StackedLSTM(input_size=config.emb_size,
hidden_size=config.decoder_hidden_size,
num_layers=config.num_layers,
dropout=config.dropout)
self.p_gen_weight = nn.Linear(
config.emb_size + config.decoder_hidden_size * 3, 1)
self.linear = nn.Linear(config.decoder_hidden_size, vocab_size)
self.output_merge = nn.Linear(2 * config.decoder_hidden_size,
config.decoder_hidden_size)
if hasattr(config, 'att_act'):
activation = config.att_act
print('use attention activation %s' % activation)
else:
activation = None
self.attention = models.global_attention(config.decoder_hidden_size,
activation)
self.memory_attention = models.global_attention(
config.decoder_hidden_size, activation)
self.softmax = nn.Softmax(-1)
self.hidden_size = config.decoder_hidden_size
self.vocab_size = vocab_size
self.dropout = nn.Dropout(config.dropout)
self.config = config
def __init__(self, config, vocab_size, embedding=None, score_fn=None):
super(rnn_decoder, self).__init__()
if embedding is not None:
self.embedding = embedding
else:
self.embedding = nn.Embedding(vocab_size, config.emb_size)
self.rnn = StackedLSTM(input_size=config.emb_size,
hidden_size=config.hidden_size,
num_layers=config.num_layers,
dropout=config.dropout)
self.score_fn = score_fn
self.linear = nn.Linear(config.hidden_size, vocab_size)
if hasattr(config, 'att_act'):
activation = config.att_act
print('use attention activation %s' % activation)
else:
activation = None
self.attention = models.global_attention(config.hidden_size,
activation)
self.hidden_size = config.hidden_size
self.dropout = nn.Dropout(config.dropout)
self.config = config
def __init__(self, config, vocab_size, embedding=None, score_fn=None):
super(rnn_decoder, self).__init__()
if embedding is not None:
self.embedding = embedding
else:
self.embedding = nn.Embedding(vocab_size, config.emb_size)
self.rnn = StackedLSTM(input_size=config.emb_size, hidden_size=config.decoder_hidden_size,
num_layers=config.num_layers, dropout=config.dropout)
self.score_fn = score_fn
if self.score_fn.startswith('general'):
self.linear = nn.Linear(config.decoder_hidden_size, config.emb_size)
elif score_fn.startswith('concat'):
self.linear_query = nn.Linear(config.decoder_hidden_size, config.decoder_hidden_size)
self.linear_weight = nn.Linear(config.emb_size, config.decoder_hidden_size)
self.linear_v = nn.Linear(config.decoder_hidden_size, 1)
elif not self.score_fn.startswith('dot'):
self.linear = nn.Linear(config.decoder_hidden_size, vocab_size)
if hasattr(config, 'att_act'):
activation = config.att_act
print('use attention activation %s' % activation)
else:
activation = None
self.attention = models.global_attention(config.decoder_hidden_size, activation)
self.hidden_size = config.decoder_hidden_size
self.dropout = nn.Dropout(config.dropout)
self.config = config
if self.config.global_emb:
self.gated1 = nn.Linear(config.emb_size, config.emb_size)
self.gated2 = nn.Linear(config.emb_size, config.emb_size)
def __init__(self,
config,
src_vocab_size,
tgt_vocab_size,
slot_embedding=None,
value_embedding=None,
score_fn=None):
super(rnn_decoder, self).__init__()
self.slot_embedding = slot_embedding
self.vocab_size = tgt_vocab_size
self.rnn = StackedLSTM(input_size=config.emb_size,
hidden_size=config.decoder_hidden_size,
num_layers=config.num_layers,
dropout=config.dropout)
for param in self.rnn.parameters():
if len(param.shape) >= 2:
init.orthogonal_(param.data)
else:
init.constant_(param.data, 0)
self.score_fn = score_fn
self.slot_linear = nn.Linear(config.decoder_hidden_size,
config.emb_size)
init.kaiming_normal_(self.slot_linear.weight.data)
init.constant_(self.slot_linear.bias.data, 0)
activation = None
self.attention = models.global_attention(config.decoder_hidden_size,
activation)
#self.user_attention = models.global_attention(config.decoder_hidden_size, activation)
#self.sys_attention = models.global_attention(config.decoder_hidden_size, activation)
self.hidden_size = config.decoder_hidden_size
self.config = config
self.linear_out = nn.Linear(4 * self.hidden_size, self.hidden_size)
init.kaiming_normal_(self.linear_out.weight.data)
init.constant_(self.linear_out.bias.data, 0)
self.re1 = nn.ReLU()
# self.dp1=nn.Dropout(0.1)
self.linear_slot = nn.Linear(self.hidden_size, self.hidden_size)
init.kaiming_normal_(self.linear_slot.weight.data)
init.constant_(self.linear_slot.bias.data, 0)
self.re2 = nn.ReLU()
# self.dp2=nn.Dropout(0.1)
self.linear3 = nn.Linear(self.hidden_size, self.hidden_size)
init.kaiming_normal_(self.linear3.weight.data)
init.constant_(self.linear3.bias.data, 0)
self.re3 = nn.ReLU()
self.sigmoid = nn.Sigmoid()
self.log_softmax = nn.LogSoftmax(dim=1)
self.linear4 = nn.Linear(self.hidden_size, self.hidden_size)
init.kaiming_normal_(self.linear4.weight.data)
init.constant_(self.linear4.bias.data, 0)
self.re4 = nn.ReLU()
self.dropout = nn.Dropout(0.5) #config.dropout)
def __init__(self, config, vocab_size, embedding=None, score_fn=None):
super(rnn_decoder, self).__init__()
if embedding is not None:
self.embedding = embedding
else:
self.embedding = nn.Embedding(vocab_size, config.emb_size)
self.rnn = StackedLSTM(input_size=config.emb_size, hidden_size=config.decoder_hidden_size,
num_layers=config.num_layers, dropout=config.dropout)
self.score_fn = score_fn
if self.score_fn.startswith('general'):
self.linear = nn.Linear(config.decoder_hidden_size, config.emb_size)
elif score_fn.startswith('concat'):
self.linear_query = nn.Linear(config.decoder_hidden_size, config.decoder_hidden_size)
self.linear_weight = nn.Linear(config.emb_size, config.decoder_hidden_size)
self.linear_v = nn.Linear(config.decoder_hidden_size, 1)
elif score_fn == 'hinge_margin_loss':
self.toEmbLinear = nn.Linear(config.decoder_hidden_size, config.emb_size)
if config.global_emb:
self.linear = nn.Linear(config.decoder_hidden_size, vocab_size) # for global_embedding
elif score_fn == 'hybrid':
# Question： 有空搞清楚这里config的形式
# 转换到Emedding空间
self.toEmbLinear = nn.Linear(config.decoder_hidden_size, config.emb_size)
self.toCatLinear = nn.Linear(config['emb_size'], vocab_size)
self.grl = models.GradReverse(config['GRL_fraction'])
self.activation = nn.Tanh()
if config.global_emb:
self.linear = nn.Linear(config.decoder_hidden_size, vocab_size)
elif not self.score_fn.startswith('dot'):
self.linear = nn.Linear(config.decoder_hidden_size, vocab_size)
if hasattr(config, 'att_act'):
activation = config.att_act
print('use attention activation %s' % activation)
else:
activation = None
self.attention = models.global_attention(config.decoder_hidden_size, activation)
self.hidden_size = config.decoder_hidden_size
self.dropout = nn.Dropout(config.dropout)
self.config = config
if self.config.global_emb:
self.gated1 = nn.Linear(config.emb_size, config.emb_size) # 论文公式(13）的两个W矩阵
self.gated2 = nn.Linear(config.emb_size, config.emb_size)
def __init__(self, config, vocab_size, embedding=None, score_fn=None):
super(rnn_decoder, self).__init__()
if embedding is not None:
self.embedding = embedding
else:
self.embedding = nn.Embedding(vocab_size, config.emb_size)
self.rnn = StackedLSTM(input_size=config.emb_size,
hidden_size=config.hidden_size,
num_layers=config.num_layers,
dropout=config.dropout)
if score_fn.startswith('general'):
self.linear = nn.Linear(config.hidden_size, config.emb_size)
if score_fn.endswith('not'):
self.score_fn = lambda x: torch.matmul(
self.linear(x), Variable(self.embedding.weight.t().data))
else:
self.score_fn = lambda x: torch.matmul(
self.linear(x), self.embedding.weight.t())
elif score_fn.startswith('dot'):
if score_fn.endswith('not'):
self.score_fn = lambda x: torch.matmul(
x, Variable(self.embedding.weight.t().data))
else:
self.score_fn = lambda x: torch.matmul(
x, self.embedding.weight.t())
else:
self.score_fn = nn.Linear(config.hidden_size, vocab_size)
if hasattr(config, 'att_act'):
activation = config.att_act
print('use attention activation %s' % activation)
else:
activation = None
self.attention = models.global_attention(config.hidden_size,
activation)
self.hidden_size = config.hidden_size
self.dropout = nn.Dropout(config.dropout)
self.log_softmax = nn.LogSoftmax()
self.config = config