attentions.py 文件源码

def __call__(self, S, h):
    batch_size, src_len, hidden_size = S.data.shape
    S = self.inner_weight(F.reshape(S, (batch_size * src_len, hidden_size)))
    S = F.reshape(S, (batch_size, src_len, hidden_size))
    a = F.softmax(F.squeeze(F.batch_matmul(S, h), axis = 2))
    return a

# MLP layer, as of Bahdanau+ 15