def __init__(self,
dir_st,
vocab,
save=False,
dropout=0.25,
fixed_emb=False):
super(DropUniSkip, self).__init__(dir_st, vocab, save, dropout,
fixed_emb)
# Modules
self.seq_drop_x = SequentialDropout(p=self.dropout)
self.seq_drop_h = SequentialDropout(p=self.dropout)
def set_dropout(self, dropout):
self.dropout = dropout
self.drop_ir = SequentialDropout(p=dropout)
self.drop_ii = SequentialDropout(p=dropout)
self.drop_in = SequentialDropout(p=dropout)
self.drop_hr = SequentialDropout(p=dropout)
self.drop_hi = SequentialDropout(p=dropout)
self.drop_hn = SequentialDropout(p=dropout)
class BayesianGRUCell(AbstractGRUCell):
def __init__(self,
input_size,
hidden_size,
bias_ih=True,
bias_hh=False,
dropout=0.25):
super(BayesianGRUCell, self).__init__(input_size, hidden_size, bias_ih,
bias_hh)
self.set_dropout(dropout)
def set_dropout(self, dropout):
self.dropout = dropout
self.drop_ir = SequentialDropout(p=dropout)
self.drop_ii = SequentialDropout(p=dropout)
self.drop_in = SequentialDropout(p=dropout)
self.drop_hr = SequentialDropout(p=dropout)
self.drop_hi = SequentialDropout(p=dropout)
self.drop_hn = SequentialDropout(p=dropout)
def end_of_sequence(self):
self.drop_ir.end_of_sequence()
self.drop_ii.end_of_sequence()
self.drop_in.end_of_sequence()
self.drop_hr.end_of_sequence()
self.drop_hi.end_of_sequence()
self.drop_hn.end_of_sequence()
def forward(self, x, hx=None):
if hx is None:
hx = Variable(x.data.new().resize_(
(x.size(0), self.hidden_size)).fill_(0))
x_ir = self.drop_ir(x)
x_ii = self.drop_ii(x)
x_in = self.drop_in(x)
x_hr = self.drop_hr(hx)
x_hi = self.drop_hi(hx)
x_hn = self.drop_hn(hx)
r = F.sigmoid(self.weight_ir(x_ir) + self.weight_hr(x_hr))
i = F.sigmoid(self.weight_ii(x_ii) + self.weight_hi(x_hi))
n = F.tanh(self.weight_in(x_in) + r * self.weight_hn(x_hn))
hx = (1 - i) * n + i * hx
return hx