使用Pytorch实现two-head(多输出)模型的操作

def forward(self, x):
        #三层的表示层
        x = F.elu(self.fcR1(x))
        x = F.elu(self.fcR2(x))
        x = F.elu(self.fcR3(x))
		#two-head，两个head分别进行输出
        y0 = F.elu(self.fcH01(x))
        y0 = F.elu(self.fcH02(y0))
        y0 = F.elu(self.fcH03(y0))
        y1 = F.elu(self.fcH11(x))
        y1 = F.elu(self.fcH12(y1))
        y1 = F.elu(self.fcH13(y1))
        return y0, y1

    f_out_y0, _ = net(x0)
            _, f_out_y1 = net(x1)
            #实例化损失函数
            criterion0 = Loss()
            criterion1 = Loss()
            loss0 = criterion0(f_y0, f_out_y0, w0)
            loss1 = criterion1(f_y1, f_out_y1, w1)
            print(loss0.item(), loss1.item())
            #对网络参数进行初始化
            optimizer.zero_grad()
            loss0.backward()
            loss1.backward()
            #对网络的参数进行更新
            optimizer.step()

import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.distributed as dist
import torch.utils.data as data_utils
class Net(nn.Module):
    def __init__(self, n_input, n_hidden, n_output):
        super(Net, self).__init__()
        self.hidden1 = nn.Linear(n_input, n_hidden)
        self.hidden2 = nn.Linear(n_hidden, n_hidden)
        self.predict1 = nn.Linear(n_hidden*2, n_output)
        self.predict2 = nn.Linear(n_hidden*2, n_output)
    def forward(self, input1, input2): # 多输入！！！
        out01 = self.hidden1(input1)
        out02 = torch.relu(out01)
        out03 = self.hidden2(out02)
        out04 = torch.sigmoid(out03)
        out11 = self.hidden1(input2)
        out12 = torch.relu(out11)
        out13 = self.hidden2(out12)
        out14 = torch.sigmoid(out13)
        out = torch.cat((out04, out14), dim=1) # 模型层拼合！！！当然你的模型中可能不需要~
 
        out1 = self.predict1(out)
        out2 = self.predict2(out)
        return out1, out2 # 多输出！！！
net = Net(1, 20, 1)
x1 = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1) # 请不要关心这里，随便弄一个数据，为了说明问题而已
y1 = x1.pow(3)+0.1*torch.randn(x1.size())
x2 = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1)
y2 = x2.pow(3)+0.1*torch.randn(x2.size())
x1, y1 = (Variable(x1), Variable(y1))
x2, y2 = (Variable(x2), Variable(y2))
optimizer = torch.optim.SGD(net.parameters(), lr=0.1)
loss_func = torch.nn.MSELoss()
for t in range(5000):
    prediction1, prediction2 = net(x1, x2)
    loss1 = loss_func(prediction1, y1)
    loss2 = loss_func(prediction2, y2)
    loss = loss1 + loss2 # 重点！
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    if t % 100 == 0:
       print('Loss1 = %.4f' % loss1.data,'Loss2 = %.4f' % loss2.data,)