gradients的一些理解

Posted Joyce_song94

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了gradients的一些理解相关的知识,希望对你有一定的参考价值。

Each variable has a [.grad_fn] attribute that references a Function that has created the Variable(except for Variables created by the user their grad_fn is None).

out.backward()=out.backward(torch.Tensor([1.0]))

 1 import torch
 2 from torch.autograd import Variable
 3 
 4 # part 1
 5 x = Variable(torch.ones(2, 2), requires_grad=True)
 6 y = x + 2
 7 print(y.grad_fn)    # <torch.autograd.function.AddConstantBackward object at 0x000001D8A156E048>
 8 print(y.grad)       # None
 9 z = y * y * 3
10 out = z.mean()
11 out.backward()
12 print(out.grad)     # None
13 print(y.grad)       # None
14 print(x.grad)       # d(out)/dx
15 ‘‘‘
16 Variable containing:
17  4.5000  4.5000
18  4.5000  4.5000
19 [torch.FloatTensor of size 2x2]
20 ‘‘‘
21 print(x.grad_fn)    # None
22 print(x.grad_output)
23 
24 
25 # part 2
26 x = torch.randn(3)
27 x = Variable(x, requires_grad=True)
28 y = x * 2
29 # print(type(y))          # <class ‘torch.autograd.variable.Variable‘>
30 # print(type(y.data))     # <class ‘torch.FloatTensor‘>
31 # print(y.data.norm())    # 4.076032686856067
32 while y.data.norm() < 1000:
33     y = y * 2
34 
35 # print(x.grad)           # None
36 gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
37 # print(y)  # Variable containing: 377.3516 997.8206 11.2558 [torch.FloatTensor of size 3]
38 y.backward(gradients)
39 # y.backward()            # y.backward()=y.backward(torch.Tensor([1.0]))
40 # RuntimeError: grad can be implicitly created only for scalar outputs
41 
42 print(x.grad)
43 # print(x.grad_fn)          # None
44 # print(x.grad_output)      # AttributeError: ‘Variable‘ object has no attribute ‘grad_output‘

 

以上是关于gradients的一些理解的主要内容,如果未能解决你的问题,请参考以下文章

机器学习数学基础- gradient descent算法(上)

深度学习100问之深入理解Vanishing/Exploding Gradient(梯度消失/爆炸)

深入梯度下降(Gradient Descent)算法

深入理解 CSS linear-gradient

翻译:如何理解梯度下降算法Gradient Descent algorithm

金蝶handler中 collection 代码片段理解