RNN复现

import torch
import torch.nn as nn
#RNN输入维度 [seq_len,batch_size,input_dim]
x_input=torch.randn(2,3,10)
#torch.nn.RNN(input_size,hidden_size,num_layers,batch_first)
#input_size 输入维度 hidden_size隐藏层维度 num_layers隐藏层层数 
#batch_first=True时x：[batch, seq_len, input_size]，h0：[batch, num_layers, hidden_size]
#batch_first=False时x：[seq_len, batch, input_size]，h0：[num_layers, batch, hidden_size]

class RNN(nn.Module):
    def __init__(self,input_size,hidden_size,batch_first=False):
        super(RNN,self).__init__()
        self.rnn_cell=nn.RNNCell(input_size,hidden_size)        
        self.batch_first=batch_first
        self.hidden_size=hidden_size        
    #权重初始化 初始隐藏层状态
    def _initialize_hidden(self,batch_size):
        return torch.zeros((batch_size,self.hidden_size))     
    #如果初始隐藏状态未设定 则initial_hidden=None,调用_initialize_hidden    
    def forward(self,inputs,initial_hidden=None):
        if self.batch_first:
            batch_size,seq_size,feat_size=inputs.size()
            #将batch_size换至第二个位置
            inputs=inputs.permute(1,0,2)
        else:
            seq_size,batch_size,feat_size=inputs.size()       
    #记录每一个隐藏层状态的值
        hiddens=[]        
        if initial_hidden is None:
            initial_hidden=self._initialize_hidden(batch_size)
            initial_hidden=initial_hidden.to(inputs.device)
        #初始化开始时刻的隐藏状态    
        hidden_t=initial_hidden        
        for t in range(seq_size):
            #当前隐藏状态由当前输入和上一时刻的隐藏状态共同决定
            hidden_t=self.rnn_cell(inputs[t],hidden_t)
            #将当前状态加入至隐藏状态列表中
            hiddens.append(hidden_t)        
        hiddens=torch.stack(hiddens)
        if self.batch_first:
            #将batch_size换回至第一个位置，因为在上面已经将batch_size换至第二个位置了，所以此处需要换回来
            hiddens=hiddens.permute(1,0,2)            
        return hiddens
model=RNN(input_size=10,hidden_size=15,batch_first=True)
outputs=model(x_input)
print(outputs)
# print(outputs.shape)