1. requires_grad

如果需要为张量计算所需的梯度，那么我们就需要对张量设置requires_grad=True;张量创建的时候默认requires_grad=False

如果不设置requires_grad=True,后续计算梯度的时候就会报错
（1）requires_grad=False&默认设置

import torch
from torch import nn

# 创建一个输入x，默认设置
x = torch.ones(5)
# y = 2*x**2
y = 2*torch.dot(x,x)
# y 进行梯度返传
y.backward()
# 打印x的梯度，即x.grad
print(f"x.grad={x.grad}")

结果

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

（2）requires_grad=False

import torch
from torch import nn

x_false = torch.ones(5, requires_grad=False)
y_false = 2 * torch.dot(x_false, x_false)
y_false.backward()
print(f"x_false.grad={x_false.grad}")

结果

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

（3）requires_grad=True

import torch
from torch import nn


x_true = torch.ones(5, requires_grad=True)
y_true = 2 * torch.dot(x_true,x_true)
y_true.backward()
print(f"x_true.grad={x_true.grad}")
print(f"4*x_true={4*x_true}")
print(f"x_true.grad==4*x_true={x_true.grad==4*x_true}")

结果：

x_true.grad=tensor([4., 4., 4., 4., 4.])
4*x_true=tensor([4., 4., 4., 4., 4.], grad_fn=<MulBackward0>)
x_true.grad==4*x_true=tensor([True, True, True, True, True])

2. grad_fn，grad

grad:表示当执行完y.backward()后，可以通过x.grad计算x变量的梯度
grad_fn是用来记录变量是怎么来的，记录图节点的方式，为了后续反向传播做准备
$z=2*x^2+6$
由上述公式可得：

x：最底层的生物，牛马如我;故x.grad_fn=None
$y=2*x^2$ :来源于乘法，故y.grad_fn = MulBackward
$z = y + 6$ :来源于加法，故z.grad_fn = AddBackward

x_true = torch.ones(5, requires_grad=True)
y_true = 2 * torch.dot(x_true, x_true)
z_true = y_true + 6
z_true.backward()
print(f"x_true.grad={x_true.grad}")
print(f"x_true.grad_fn={x_true.grad_fn}")
print(f"y_true.grad_fn={y_true.grad_fn}")
print(f"z_true.grad_fn={z_true.grad_fn}")

结果：

x_true.grad=tensor([4., 4., 4., 4., 4.])
x_true.grad_fn=None
y_true.grad_fn=<MulBackward0 object at 0x00000180E0DF3550>
z_true.grad_fn=<AddBackward0 object at 0x00000180E0DF3550>

3. with torch.no_grad()

torch.no_grad
禁用梯度计算的上下文管理器;当您确定不会调用Tensor.backward()时，禁用梯度计算对推断很有用。它将减少原本需要requires_grad=True的计算的内存消耗
有两种方式设置：

with torch.no_grad()
@torch.no_grad()

# 定义一个张量x
x = torch.tensor([1.0], requires_grad=True)

# 用with torch.no_grad()防止计算梯度
with torch.no_grad():
	y = x * 2

print(f"y.requires_grad={y.requires_grad}")
#输出结果： y.requires_grad=False
# 用@torch.no_grad()

@torch.no_grad()
def doubler(x):
	return x * 2


z = doubler(x)
print(f"z.requires_grad={z.requires_grad}")
# z.requires_grad=False

4. tensor.detach()

返回一个新的张量，从当前图分离出来。

import torch
from torch import nn

x = torch.ones(5)
# destination
y = torch.zeros(3)
w = torch.randn((5, 3), requires_grad=True)
b = torch.randn(3, requires_grad=True)
z = torch.matmul(x, w) + b
loss = torch.nn.functional.binary_cross_entropy_with_logits(z,y)

print(f"z.grad_fn={z.grad_fn}")
print(f"loss.grad_fn={loss.grad_fn}")
loss.backward()
print(f"w.grad={w.grad}")
print(f"b.grad={b.grad}")
print(f"z.requires_grad={z.requires_grad}")
# with torch.no_grad():

# 新建一个张量z_dect
z_dect = z.detach()
print(f"after_detach:z.requires_grad={z_dect.requires_grad}")