Hexo

问题描述

`requires_grad=True`

等价于requires_grad=a, a为任意非0整数,不能为浮点数
浮点数会报错: TypeError: integer argument expected, got float

测试代码

import torch
from torch.autograd import Variable

tensor = torch.ones(1)
variable = Variable(tensor, requires_grad=True)
print(tensor)
print(variable)
print("type1: ", type(tensor), type(variable))
print(tensor.data)
print(variable.data)
print("type2: ", type(tensor.data), type(variable.data))
print(tensor.data.numpy())
print(variable.data.numpy())
print("type3: ", type(tensor.data.numpy()), type(variable.data.numpy()))
print(tensor.numpy())
print(variable.numpy())
print("type4: ", type(tensor.numpy()), type(variable.numpy()))

# Output:
tensor([1.])
tensor([1.], requires_grad=True)
('type1: ', <class 'torch.Tensor'>, <class 'torch.Tensor'>)
tensor([1.])
tensor([1.])
('type2: ', <class 'torch.Tensor'>, <class 'torch.Tensor'>)
[1.]
[1.]
('type3: ', <type 'numpy.ndarray'>, <type 'numpy.ndarray'>)
[1.]
Traceback (most recent call last):
  File "/home/sanye/JupyterWorkspace/test.py", line 16, in <module>
    print(variable.numpy())
RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

从上面的测试用例可以看出:
- Variable和Tensor在判断类型时都是torch.Tensor
  - type(tensor) == type(variable) == torch.Tensor
- 几乎所有操作都相同
  - tensor.data == variable.data
  - tensor.data.numpy() == varible.data.numpy()
- 直接输出变量结果不相同
  - tensor输出时没有requires_grad=True
  - variable输出时有requires_grad=True
- variable不能直接调用函数variable.numpy(),会报异常
  - 异常描述为: 当前Variable变量要求requires grad,也就是requires_grad属性为真时,变量不能直接使用

`requires_grad=False`

等价于requires_grad=0
不等价于requires_grad=None, None会报错: TypeError: an integer is required

测试代码:

import torch
from torch.autograd import Variable

tensor = torch.ones(1)
variable = Variable(tensor, requires_grad=False)
print(tensor)
print(variable)
print("type1: ", type(tensor), type(variable))
print(tensor.data)
print(variable.data)
print("type2: ", type(tensor.data), type(variable.data))
print(tensor.data.numpy())
print(variable.data.numpy())
print("type3: ", type(tensor.data.numpy()), type(variable.data.numpy()))
print(tensor.numpy())
print(variable.numpy())
print("type4: ", type(tensor.numpy()), type(variable.numpy()))

# Output:
tensor([1.])
tensor([1.])
('type1: ', <class 'torch.Tensor'>, <class 'torch.Tensor'>)
tensor([1.])
tensor([1.])
('type2: ', <class 'torch.Tensor'>, <class 'torch.Tensor'>)
[1.]
[1.]
('type3: ', <type 'numpy.ndarray'>, <type 'numpy.ndarray'>)
[1.]
[1.]
('type4: ', <type 'numpy.ndarray'>, <type 'numpy.ndarray'>)

从上面的测试用例可以看出:
- 当variable变量的requires_grad=False时,variable完全退化为tensor
  - 直接输出变量时没有requires_grad=False属性
  - 可以直接使用variable.numpy()函数

`Variable`的三种等价定义

下面三种定义的Variable类型变量varible等价

requires_grad=False

1	variable = Variable(tensor, requires_grad=False)

没有requires_grad参数
1
variable = Variable(tensor)

requires_grad=True,然后variable = variable.detach()

1 2	variable = Variable(tensor, requires_grad=True) variable = variable.detach()

上面三种定义都等价于原始的tensor
- 这里的等价并未经过详细测试,但是至少以下方面等价:
  - 自身类型相同type, 类型为torch.Tensor
  - 可以调用属性.data,类型为torch.Tensor
  - 可以调用.grad,只不过都为None
  - 直接输出对象完全相同,都不包含requires_grad=True属性
  - 可以调用相同的函数.numpy(), 类型为numpy.ndarray
  - 可以调用相同的函数.data.numpy(), 类型为numpy.ndarray