Pytorch 自动微分注意点讲解

backward()

backward()函数是pytorch框架实现自动微分的关键函数,一般通过loss.backward()调用,这里的loss一般是标量张量

python">import numpy as np
import torch
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True,device=device)
print(data1)
y = data1.pow(2)+100
print('y =',y)
loss = y.mean()
print('loss =',loss)
loss.backward()
print("data1's grad =",data1.grad)
# mps
# tensor([[1., 9., 0.]], device='mps:0', requires_grad=True)
# y = tensor([[101., 181., 100.]], device='mps:0', grad_fn=<AddBackward0>)
# loss = tensor(127.3333, device='mps:0', grad_fn=<MeanBackward0>)
# data1's grad = tensor([[0.6667, 6.0000, 0.0000]], device='mps:0')

可以看到这里模拟了一个函数计算,mean()模拟了损失计算,目的是将食粮张量转为标量张量

在对损失loss进行了反向传播后,叶子节点data1便有了grad属性.也就是2*data1

设备迁移注意点

python">import numpy as np
import torch
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True).to(device)
print(data1)
y = data1.pow(2)+100
print('y =',y)
loss = y.mean()
print('loss =',loss)
loss.backward()
print("data1's grad =",data1.grad)
# mps
# tensor([[0., 9., 0.]], device='mps:0', grad_fn=<ToCopyBackward0>)
# y = tensor([[100., 181., 100.]], device='mps:0', grad_fn=<AddBackward0>)
# loss = tensor(127., device='mps:0', grad_fn=<MeanBackward0>)
# /Users/jinhouji/PycharmProjects/pythonProject/lesson01.py:13: UserWarning:
# 
# The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at /Users/runner/work/pytorch/pytorch/pytorch/build/aten/src/ATen/core/TensorBody.h:494.)
# 
# data1's grad = None

这里在创建完张量进行to(device)后,会出现打印梯度为None的情况,这是由于在张量创立后进行设备转移操作会导致grad_fn(也就是函数操作记录)的丢失,所以这个时候可以通过detach()+requires_grad_(True)函数来重新建立叶子节点

detach()

detach()函数可以建立一个与原张量共享内存但不进行梯度计算的全新张量

clone()

clone()函数可以拷贝一个和原张量具有相同计算图和张量值的全新张量

requires_grad_()

requires_grad_()函数可以将张量的梯度计算权限打开

is_leaf()

is_leaf()函数用于判断张量是否为叶子节点,返回布尔值

综上,我们可以尝试去重建叶子节点

python">import numpy as np
import torch
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True).to(device).detach().requires_grad_(True)
print(data1)
print(data1.is_leaf)
y = data1.pow(2)+100
print('y =',y)
loss = y.mean()
print('loss =',loss)
loss.backward()
print("data1's grad =",data1.grad)
# mps
# tensor([[7., 8., 8.]], device='mps:0', requires_grad=True)
# True
# y = tensor([[149., 164., 164.]], device='mps:0', grad_fn=<AddBackward0>)
# loss = tensor(159., device='mps:0', grad_fn=<MeanBackward0>)
# data1's grad = tensor([[4.6667, 5.3333, 5.3333]], device='mps:0')

以上为使用了detach()和requires_grad()的方法,所以注意要将非叶子节点拆分为叶子节点的方法就是以上所示的流程

暂停梯度计算方法

一般在做模型推理和评估的时候需要暂停梯度计算,以下列举三种停止梯度计算的方法

with torch.no_grad():

python">import numpy as np
import torch
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True).to(device).clone()
print(data1)
print(data1.is_leaf)
with torch.no_grad():y = data1.pow(2)+100
print(y.requires_grad)
# mps
# tensor([[0., 5., 5.]], device='mps:0', grad_fn=<CloneBackward0>)
# False
# False

with代码块下的语句涉及张量生成的代码,都不会进行梯度计算

@torch.no_grad()

@torch.no_grad()装饰器装饰的函数内部若进行张量生成则不会进行梯度计算

python">import numpy as np
import torch
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True).to(device).clone()
print(data1)
print(data1.is_leaf)
@torch.no_grad()
def func11(data1):return data1.pow(2)+100
print(func11(data1).requires_grad)
# mps
# tensor([[3., 4., 4.]], device='mps:0', grad_fn=<CloneBackward0>)
# False
# False

torch.set_grad_enabled()

torch.set_grad_enabled()函数可以通过设置参数为False来关闭梯度计算,如需开启梯度计算则需要重新调用函数设置参数为True

python">import numpy as np
import torch
from torch import set_grad_enableddevice = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
print(device
)
data1 = torch.randint(0,10,(1,3),dtype=torch.float,requires_grad=True).to(device).clone()
print(data1)
print(data1.is_leaf)
torch.set_grad_enabled(False)
y=data1.pow(2)+100
print(y.requires_grad)
set_grad_enabled(True)
y=data1.pow(2)+100
print(y.requires_grad)
# mps
# tensor([[7., 8., 6.]], device='mps:0', grad_fn=<CloneBackward0>)
# False
# False
# True