RNN

RNN循环神经网络，主要用于处理和预测序列数据
RNN和LSTM - 图1
CLASS torch.nn.``RNN(*args, **kwargs)
image.png

  1. class torch.nn.RNN(*args, **kwargs)
  2. Parameters:
  3.     input_size:
  4.     hidden_size:
  5.       num_layers:     default:1
  6.     nonlinearity:    可以选择激活函数'tanh'或者'relu'，default:tanh
  7.     bias:            如果是False的话，不会用b_ih和b_hh，default:True
  8.     batch_first:    如果是True，输入的shape是(batch,seq_len,feature)，default:False,(seq_len,batch,feature)
  9.     drop_out:       default:0
  10.     bidirectional:  如果是true，变成双向，default:False
  12. Inputs:input,h_0
  13.     input的shape是(seq_len,batch,input_size),
  14.     h_0的shape是(num_layers * num_directions,batch,hidden_size)，如果bidirectional是True，
  15.     num_directions是2，否则为1
  16. Outputs:output,h_n
  17.     output的shape是(seq_len,batch,num_directions*hidden_size)，是每个h_t的值
  18.     h_n的shape是(num_layers*num_directions,batch,hidden_size)，是h_t，t=seq_len时，h_t的值
  19.     output[-1,:,:]=h_n,batch_first为False

下面是torch官网自己构建的简单的RNN模型

  1. class RNN(nn.Module):
  3.     # you can also accept arguments in your model constructor
  4.     def __init__(self, data_size, hidden_size, output_size):
  5.         super(RNN, self).__init__()
  7.         self.hidden_size = hidden_size
  8.         input_size = data_size + hidden_size
  10.         self.i2h = nn.Linear(input_size, hidden_size)
  11.         self.h2o = nn.Linear(hidden_size, output_size)
  13.     def forward(self, data, last_hidden):
  14.         input = torch.cat((data, last_hidden), 1)
  15.         hidden = self.i2h(input)
  16.         output = self.h2o(hidden)
  17.         return hidden, output

LSTM

image.png

  1. class torch.nn.LSTM(*args, **kwargs)
  2. Parameters(参数同RNN):
  3.     input_size:
  4.     hidden_size:
  5.       num_layers:     
  6.     bias:            
  7.     batch_first:    
  8.     drop_out:       
  9.     bidirectional:  
  10. Input:input,(h_0,c_0)
  11.     input的shape是(seq_len, batch, input_size)
  12.     h_0的shape是(num_layers * num_directions, batch, hidden_size)