dyer_lstm.py 文件源码

def __call__(self, xs):
        """
        xs: (batchsize, hidden_dim)
        """

        if self.h is not None:
            h = self.h
            c = self.c
        else:
            xp = chainer.cuda.get_array_module(xs.data)
            batchsize = xs.shape[0]
            h = Variable(xp.zeros((batchsize, self.outsize), 'f'), volatile='AUTO')
            c = Variable(xp.zeros((batchsize, self.outsize), 'f'), volatile='AUTO')

        in_gate = F.sigmoid(self.linear_in(F.concat([xs, h, c])))
        new_in = F.tanh(self.linear_c(F.concat([xs, h])))
        self.c = in_gate * new_in + (1. - in_gate) * c
        out_gate = F.sigmoid(self.linear_out(F.concat([xs, h, self.c])))
        self.h = F.tanh(self.c) * out_gate
        return self.h