def __init__(self,
vocab_size,
embedding_size=300,
hidden_size=512,
num_layers=2,
dropout=0.3,
shared_weight=True,
**kwargs):
super().__init__()
self.embedding = Embeddings(num_embeddings=vocab_size,
embedding_dim=embedding_size,
dropout=dropout)
self.rnn = nn.LSTM(input_size=embedding_size,
hidden_size=hidden_size,
num_layers=num_layers,
bidirectional=False,
dropout=dropout,
batch_first=True)
# 输出层
self.output = nn.Sequential(nn.Linear(hidden_size, hidden_size),
nn.Tanh(),
nn.Linear(hidden_size, embedding_size))
# 投影层
self.proj = nn.Linear(embedding_size, vocab_size, bias=False)
if shared_weight:
self.proj.weight = self.embedding.embeddings.weight
else:
my_init.default_init(self.proj.weight)
def _build_bridge(self):
if self.bridge_type == "mlp":
self.linear_bridge = nn.Linear(in_features=self.context_size, out_features=self.hidden_size)
my_init.default_init(self.linear_bridge.weight)
elif self.bridge_type == "zero":
pass
else:
raise ValueError("Unknown bridge type {0}".format(self.bridge_type))
def __init__(self,
d_model, n_head,
feature_size=1024,
hidden_size=512,
dropout=0.0,
**kwargs
):
super(QE_ATTENTION, self).__init__()
self.ctx_attn = MultiHeadedAttention(head_count=n_head, model_dim=d_model, dropout=dropout,
dim_per_head=None)
# Use PAD
self.gru = RNN(type="gru", batch_first=True, input_size=feature_size, hidden_size=hidden_size,
bidirectional=True)
self.lstm = RNN(type="lstm", batch_first=True, input_size=feature_size, hidden_size=hidden_size,
bidirectional=True)
self.w = nn.Linear(2 * hidden_size, 1)
my_init.default_init(self.w.weight)
self.dropout = nn.Dropout(dropout)
self.sigmoid = nn.Sigmoid()
def __init__(self,
feature_size=768,
hidden_size=512,
dropout_rate=0.1,
**kwargs):
super(QE_PAIR, self).__init__()
# Use PAD
self.gru = RNN(type="gru",
batch_first=True,
input_size=feature_size,
hidden_size=hidden_size,
bidirectional=True)
self.lstm = RNN(type="lstm",
batch_first=True,
input_size=feature_size,
hidden_size=hidden_size,
bidirectional=True)
self.lstm_src = RNN(type="lstm",
batch_first=True,
input_size=feature_size,
hidden_size=hidden_size,
bidirectional=True)
self.lstm_mt = RNN(type="lstm",
batch_first=True,
input_size=feature_size,
hidden_size=hidden_size,
bidirectional=True)
self.w = nn.Linear(2 * hidden_size, 1)
my_init.default_init(self.w.weight)
self.w_all = nn.Linear(2 * 2 * hidden_size, 1)
my_init.default_init(self.w_all.weight)
self.w_1 = nn.Linear(2 * hidden_size, 1)
my_init.default_init(self.w_1.weight)
self.w_2 = nn.Linear(2 * hidden_size, 1)
my_init.default_init(self.w_2.weight)
self.dropout = nn.Dropout(dropout_rate)
self.sigmoid = nn.Sigmoid()
def _reset_parameters(self):
my_init.default_init(self.linear_input.weight)
my_init.default_init(self.linear_hidden.weight)
my_init.default_init(self.linear_ctx.weight)
def reset_parameters(self):
init.default_init(self.linear.weight.data)
def _reset_parameters(self):
for weight in self.parameters():
my_init.default_init(weight)
def _reset_paramters(self):
my_init.default_init(self.intrinsic_predict_FFN1.weight)
my_init.default_init(self.intrinsic_predict_FFN2.weight)
my_init.default_init(self.reconstruct_action_FFN1.weight)
my_init.default_init(self.reconstruct_action_FFN2.weight)
def _reset_parameters(self):
# GRUs are initialized in itself.
self.layer_norm.reset_parameters()
my_init.default_init(self.ffn.weight)
def _reset_parameters(self):
self.LN.reset_parameters()
my_init.default_init(self.scorer_ffn.weight)
def _reset_parameters(self):
my_init.default_init(self.ctx_linear.weight)
my_init.default_init(self.input_linear.weight)
self.rephrase_LN.reset_parameters()
# RNN params are initialized by the RNN module
my_init.default_init(self.rephraser_linear_base_mu.weight)
my_init.default_init(self.rephraser_linear_base_log_sig.weight)
my_init.default_init(self.rephraser_linear_mu.weight)
my_init.default_init(self.rephraser_linear_log_sig.weight)
def _reset_parameters(self):
my_init.default_init(self.ctx_linear.weight)
my_init.default_init(self.input_linear.weight)
my_init.default_init(self.action_linear.weight)
self.critic_LN.reset_parameters()
my_init.default_init(self.critic1_linear_base.weight)
my_init.default_init(self.critic2_linear_base.weight)
my_init.default_init(self.critic1_linear.weight)
my_init.default_init(self.critic2_linear.weight)
def _reset_parameters(self):
my_init.default_init(self.ctx_linear.weight)
my_init.default_init(self.input_linear.weight)
my_init.default_init(self.attacker_linear.weight)
my_init.default_init(self.critic_linear.weight)
def _reset_parameters(self):
if self.cell_type == 'cgru':
my_init.default_init(self.linear_input.weight)
my_init.default_init(self.linear_hidden.weight)
my_init.default_init(self.linear_ctx.weight)
def _reset_parameters(self):
for weight in self.parameters():
my_init.default_init(weight)
if self.cover_size is not None:
for weight in self.gru.parameters():
my_init.rnn_init(weight)
