Pytorch专题实战——反向传播(Backpropagation)

1.前言

大体分为三步：
（1）前向传播，计算loss
（2）计算局部梯度
（3）反向传播，用链式求导法则计算梯度

2.初识如何更新梯度

import torch

x = torch.tensor(1.0)   #指定输入x
y = torch.tensor(2.0)   #指定输出y
w = torch.tensor(1.0, requires_grad=True) #初始化待求的w， requires_grad=True代表需要求导

y_predicted = w*x     #预测值
loss = (y - y_predicted)**2     #计算预测值与真实值间的loss
print(loss)

loss.backward()       #反向传播 
print(w.grad)        #打印此时的梯度

with torch.no_grad():
    w -= 0.01*w.grad    #更新梯度
    print(w)          #打印更新后的梯度
w.grad.zero_()      #梯度清零，防止梯度累加

3.手动更新梯度

import numpy as np

X = np.array([1,2,3,4], dtype=np.float32)    #输出X
Y = np.array([2,4,6,8], dtype=np.float32)    #输出Y

w = 0.0     #初始化权重

def forward(x):
    return w*x      #预测值

def loss(y, y_pred):
    return ((y_pred-y)**2).mean()    #计算loss

#Loss = 1/N*(w*x-y)**2
#dLoss/dw = 1/N*2x(w*x-y)
def gradient(x,y,y_pred):          #计算梯度
    return np.dot(2*x, y_pred-y).mean()

print(f'Prediction before training:f(5)={forward(5):.3f}')    #当权重为0是，预测值为0

learning_rate = 0.01    #学习率
n_iters = 20      #迭代次数

for epoch in range(n_iters): 
    y_pred = forward(X)    #预测值
    l = loss(Y,y_pred)     #真实值与预测值的误差
    dw = gradient(X,Y,y_pred)    #求梯度
    w -= learning_rate * dw      #更新参数
    if epoch % 2 == 0:          #每隔两个epoch更新参数
        print(f'epoch {epoch+1}: w={w:.3f},loss={l:.8f}')
print(f'Prediction after training:f(5)={forward(5):.3f}')    #运行最后的权重算f(5)

4.自动更新梯度

import torch

X = torch.tensor([1,2,3,4], dtype=torch.float32)
Y = torch.tensor([2,4,6,8], dtype=torch.float32)

w = torch.tensor(0.0, dtype=torch.float32, requires_grad=True)    #初始化权重，并设置需要求导

def forward(x):
    return w * x      #预测

def loss(y, y_pred):
    return ((y_pred-y)**2).mean()   #算loss
print(f'Prediction before training:f(5)={forward(5).item():.3f}')   #打印初始的预测

learning_rate = 0.1
n_iters = 20

for epoch in range(n_iters):
    y_pred = forward(X)
    l = loss(Y, y_pred)
    l.backward()   #等同于计算梯度
    
    with torch.no_grad():
        w -= learning_rate*w.grad     #更新参数
    w.grad.zero_()     #梯度清零，防止梯度累计
    if epoch % 2==0:
        print(f'epoch {epoch+1}: w={w.item():.3f}, loss={l.item():.8f}')
print(f'Prediction after training:f(5)={forward(5).item():.3f}')