LSTM一种优雅的参数初始化方式

class LstmRelu(nn.Module):
    @staticmethod
    def weight_init(m):
        if isinstance(m,nn.Linear):
            nn.init.xavier_normal_(m.weight)
            nn.init.constant_(m.bias,0)
        elif isinstance(m,nn.BatchNorm1d):
            nn.init.constant_(m.weight,1)
            nn.init.constant_(m.bias,0)
            
    def __init__(self,input_size,hidden_size,num_layers):
        super(LstmRelu,self).__init__()
        self.lstm = nn.LSTM(input_size,hidden_size,num_layers)
        for name, param in self.lstm.named_parameters():
            nn.init.uniform_(param,-0.1,0.1)
        self.lin_1 = nn.Linear(hidden_size,1000)
        self.lin_2 = nn.Linear(1000,1000)
        self.lin_3 = nn.Linear(1000,1000)
        self.apply(self.weight_init)
        
    def forward(self,x): # [seq_len,batch,embed],只要保证embed和input相等就行
        out, _ = self.lstm(x) # seq_len,batch,hidden 
        out = self.lin_1(out)
        out = F.relu(out)
        out = self.lin_2(out)
        out = F.relu(out)
        out = self.lin_3(out)
        out = F.relu(out)
        return out