李沐的深度学习课 课程笔记0.1 自动求导

import torch


def demo1(x: torch.Tensor):
    # tensor([0., 1., 2., 3.])
    x.requires_grad_(True)  # tensor([0., 1., 2., 3.], requires_grad=True)

    print(x)
    # 相当于 torch.arange(4.0, requires_grad=True)
    y = 2 * torch.dot(x, x)  # tensor(28., grad_fn=<MulBackward0>)
    y.backward()  # 通过调用反向传播函数来自动计算y关于x每个分量上的梯度
    print(x.grad)
    print(x.grad == 4 * x)

    # 默认情况下,PyTorch会累计梯度,所以再次计算时需要清除之前的计算值
    x.grad.zero_()
    y = 2 * torch.dot(x, 2 * x)
    y.backward()
    print(x.grad)


def f(x):
    return 3 * torch.dot(x, x) + 5


def demo2(x: torch.Tensor):
    x.requires_grad_(True)
    print(x)
    y = f(x)
    y.backward()
    print(x.grad)


if __name__ == '__main__':
    x = torch.arange(4.0)
    demo2(x)