PyTorch张量的自动微分

梯度下降法求解最优解

import torch

def f01():
    W = torch.tensor(10,requires_grad=True,dtype=torch.float)
    # 学习率
    lr = 0.001

    for i in range(5000):
        # 前向传播
        y = W ** 2 + 20
        # 防止梯度累加
        if W.grad is not None:
            W.grad.zero_()
        # 反向传播
        y.backward()
        # 梯度下降法公式
        W.data = W.data - lr * W.grad
        print(f"第{i}轮，w为{W}",(i,W))

if __name__ == '__main__':
    f01()

# 将带有梯度的张量，转换成numpy
def f02():
    w = torch.tensor([10.,10.],requires_grad=True,dtype=torch.float)
    w1 = w.detach()
    print(w.requires_grad) # True
    print(w1.requires_grad) #Flase
    print(w.data)
    print(w1.data)
    print(id(w.data))
    print(id(w1.data))

    w2 = w1.numpy()
    print(id(w2),w2,type(w2))

自动微分模块应用

def f03():
    x = torch.ones(2,5)
    y = torch.zeros(2,3)
    # 创建权重w和偏置b
    w = torch.randn(5,3,requires_grad=True)
    b = torch.randn(3,requires_grad=True)

    # 网络输出
    z = torch.matmul(x,w) + b
    loss = torch.nn.MSELoss()
    loss = loss(z,y)
    print(loss.item())
    loss.backward()
    print("w的梯度为：",w.grad)
    print("b的梯度为：",b.grad)