torch的LSTM输出问题

class Net(torch.nn.Module):
    def __init__(self):
        super(Net,self).__init__()
        self.bilstm = torch.nn.LSTM(4,10,batch_first=False,bidirectional=True)
    def forward(self,input):
        rnn_out,_ = self.bilstm(input)  #输出形状 [length:3,batch:2,hidden:20]
        encoding = torch.cat([rnn_out[0], rnn_out[-1]], dim=1)  ## 选择双向hidden的最后一个向量，并把他们拼接成一个向量
        #cat  拼接向量  dim=0是按第一个维度拼接，dim=1是按第二个维度拼接，dim=2 以此类推
        return encoding  

net = Net()
input = torch.rand(6,2,4)
input
out = net(input)
out

输出：
tensor([[[0.8778, 0.0818, 0.6875, 0.8859],
         [0.5734, 0.2775, 0.3590, 0.2174]],

        [[0.9982, 0.5425, 0.3168, 0.1968],
         [0.3750, 0.4696, 0.5498, 0.0981]],

        [[0.1177, 0.0353, 0.0145, 0.5931],
         [0.9969, 0.8779, 0.0992, 0.0971]],

        [[0.2751, 0.5499, 0.0971, 0.6711],
         [0.1580, 0.5890, 0.0929, 0.2064]],

        [[0.5682, 0.0360, 0.0856, 0.8951],
         [0.1791, 0.3881, 0.4138, 0.0593]],

        [[0.8850, 0.2704, 0.5389, 0.3144],
         [0.8112, 0.8874, 0.0174, 0.1409]]])

tensor([[-0.0343,  0.1127,  0.0279,  0.0927,  0.0789,  0.0817, -0.0883,  0.1394,
         -0.0650,  0.0198, -0.2320, -0.1571, -0.2744,  0.0421, -0.0620, -0.1478,
         -0.0336, -0.1328,  0.1629, -0.4289, -0.1029,  0.1805,  0.0289,  0.1386,
          0.1701,  0.1697, -0.1425,  0.1867, -0.0695, -0.0141, -0.1513, -0.0761,
         -0.1537,  0.0194, -0.0216, -0.0107,  0.0298, -0.1285,  0.0615, -0.1595],
        [-0.0676,  0.0644, -0.0016,  0.0671,  0.0838,  0.0723, -0.0314,  0.0668,
         -0.0258, -0.0100, -0.2025, -0.1248, -0.2316, -0.0192, -0.1111, -0.0949,
          0.0649, -0.1628,  0.0912, -0.3886, -0.1119,  0.0934,  0.0270,  0.0647,
          0.1424,  0.1345, -0.0460,  0.1188, -0.0534, -0.0672, -0.1266, -0.0497,
         -0.1637, -0.0041, -0.0519,  0.0089,  0.0486, -0.1394, -0.0221, -0.1710]],
       grad_fn=<CatBackward>)

转载于:https://www.cnblogs.com/rise0111/p/11533661.html