RNN-torch.nn.RNN

RNN-torch.nn.RNN

RNN的本质就是就是：hidden = tanh(linear([input+hidden]))

调用pytorch的nn.RNN接口，只需要明确一点，input的默认形状是[seqlength, batchsize, feature_size]， hidden的形状是[num_layers, batch_size, hidden_size]。返回两个结果：所有hidden（如果num_layers>1，就返回所有hidden[-1]）, 以及最后一个hidden。

import torch
import torch.nn as nn

input = torch.randn(5, 3, 10) # seq_length=5, batch_size=3, feature_size=10
hidden0 = torch.randn(2, 3, 20) # num_layers=2, batch_size=3, hidden_size=20

rnn = nn.RNN(input_size=10, hidden_size=20, num_layers=2)
all_hidden, last_hidden = rnn(input, hidden0) 
print(all_hidden.size()) # [5, 3, 20] [seq_length, batch_size, hidden_size]]
print(last_hidden.size()) # [2, 3, 20] [num_layers, batch_size, hidden_size]