BertLSTM的forward函数中对hidden操作写的似乎有些问题

[bluesea0]

#BertLSTM.py    line 19:
self.rnn = nn.LSTM(config.hidden_size, rnn_hidden_size, num_layers,bidirectional=bidirectional, batch_first=True, dropout=dropout)
.....

#BertLSTM.py    line 31-36:
_, (hidden, cell) = self.rnn(encoded_layers)
# outputs: [batch_size, seq_len, rnn_hidden_size * 2]
hidden = self.dropout(
torch.cat((hidden[-2, :, :], hidden[-1, :, :]), dim=1))  # 连接最后一层的双向输出

logits = self.classifier(hidden)

这里由于之前设置了batch_first=True，hidden的shape=[batch_size,num_layers*nums_directions,rnn_hidden_size]，所以连接最后一层双向输出应该是(hidden[:,-2,:],hidden[:,-1,:])吧？或者将hidden=hidden.permute([1,0,2])

[bluesea0]

明白了，设置batch_first=True只影响到output的shape，对hidden和cell的shape不会有影响。