链接：https://www.zhihu.com/question/398425328/answer/1454276131

题主已经标明了是Torch，那么没区别，backward只是积累梯度而已，哪怕同一个loss，你backward两次，就是积累两倍梯度。要使用optimizer.step之后才会根据积累的梯度更新权重。至于梯度是怎么积累的，optimizer不会管。虽然其实和题目无关，但前面的答主们很嗨的研究多任务优化问题，还是很赞的。。。跑下面的程序，打印出来的梯度是一样的：import torch

import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as Finput = torch.rand((1, 64)).float()
ground_truth1 = torch.tensor([[0]]).float()
ground_truth2 = torch.tensor([[0]]).float()class MyModel(nn.Module):def __init__(self):super(MyModel, self).__init__()self.layer1 = nn.Linear(64, 8, bias=False)self.layer2 = nn.Linear(8, 1, bias=False)def forward(self, x):out1 = self.layer1(x)out2 = self.layer2(out1)return F.sigmoid(out2), F.sigmoid(out2)model = MyModel()
model.train()optimizer = optim.SGD(model.parameters(), 0.01)
output1, output2 = model(input)loss1 = F.mse_loss(output1, ground_truth1)
loss2 = F.mse_loss(output2, ground_truth2)optimizer.zero_grad()
loss1.backward(retain_graph=True)
loss2.backward(retain_graph=True)
print(model.layer1.weight.grad[0, 0])optimizer.zero_grad()
loss = (loss1 + loss2)
loss.backward()
print(model.layer1.weight.grad[0, 0])