optimizer.zero_grad()、loss.backward()、optimizer.step()

 1、Theory

loss.backward() computes dloss/dx for every parameter x which has requires_grad=True. These are accumulated into x.grad for every parameter x. In pseudo-code:

x.grad += dloss/dx

optimizer.step updates the value of x using the gradient x.grad. For example, the SGD optimizer performs:

x += -lr * x.grad

optimizer.zero_grad() clears x.grad for every parameter x in the optimizer. It’s important to call this before loss.backward(), otherwise you’ll accumulate the gradients from multiple passes.

If you have multiple losses (loss1, loss2) you can sum them and then call backwards once:

loss3 = loss1 + loss2
loss3.backward()

2、Example 

if you have “net2” which is a pretrained network and you want to backprop the gradients of the loss of “net2” into “net1”. In pseudo-code: 

import torch
from torch import optim


def train(n_epoch):
    net1 = model1()
    net2 = model2(pretrained=True)
    optim1 = optim.SGD(net1.parameters(), lr=0.1)

    for epoch in range(n_epoch):
        for data in dataloader:
            net2.eval()
            with torch.no_grad():
                loss2 = net2(data)

            net1.train()
            optim1.zero_grad()
            loss1 = net1(data)
            total_loss = loss1 + loss2
            total_loss.backward()
            optim1.step()

参考： What does the backward() function do? - autograd - PyTorch Forums