detach()[source]
Returns a new Variable, detached from the current graph. 返回一个新的Variable，从当前图中分离下来的。
返回的Variable requires_grad=False，如果输入 volatile=True，那么返回的Variable volatile=True。
注意：
返回的Variable和原始的Variable公用同一个data tensor。in-place修改会在两个Variable上同时体现(因为它们共享data tensor)，可能会导致错误。
文章目录
- detach()[source]
- 20230816
- PyTorch的torch.detach()函数：详解与应用
- 目录
- 1. torch.detach()函数简介
- 2. torch.detach()的使用场景
- 3. torch.detach()和.requires_grad_的区别
- 4. 代码示例
- 5. 总结
- 参考资料
20230816
PyTorch的torch.detach()函数：详解与应用
PyTorch是一个广泛使用的深度学习库，其中包含各种实用的函数和方法。在本文中，我们将详细介绍torch.detach()这个函数，它返回一个新的Variable，从当前计算图中分离下来。
目录
- torch.detach()函数简介
- torch.detach()的使用场景
- torch.detach()和.requires_grad_的区别
- 代码示例
- 总结
- 技术投票
1. torch.detach()函数简介
torch.detach()函数创建了一个新的Tensor，该Tensor与原始Tensor共享数据但不需要梯度计算。换句话说，它会从当前的计算图中分离出一个新的Tensor，使得对该Tensor的进一步操作不会影响到计算图的梯度传播。
new_tensor = old_tensor.detach()2. torch.detach()的使用场景
在深度学习训练过程中，我们有时候希望某些操作不参与梯度计算。比如，在某些Reinforcement Learning算法中，策略网络的输出（即动作）常常需要从计算图中剥离开，使得在后续计算奖励或者其他指标时不会影响到策略网络的参数更新。
此外，torch.detach()也常用于保存模型的中间输出，而这些输出并不需要在反向传播中被计算。
3. torch.detach()和.requires_grad_的区别
虽然torch.detach()和.requires_grad_都能使Tensor脱离计算图，但他们的工作方式稍有不同。
-
torch.detach()返回一个新的Tensor，与原Tensor共享数据但不需要梯度计算。 -
.requires_grad_则直接更改原Tensor的属性，使其不再需要梯度计算。
在大多数情况下，torch.detach()更为安全，因为它不会修改原始Tensor的属性。
4. 代码示例
下面的代码片段演示了torch.detach()在PyTorch中的使用。
import torch
# 创建一个需要梯度计算的Tensor
x = torch.tensor([1.0, 2.0, 3.0], requires_grad=True)
y = x * 2
# 使用torch.detach()创建一个新的Tensor
z = y.detach()
print('Original Tensor: ', y)
print('Detached Tensor: ', z)在上述代码中，y是一个需要梯度计算的Tensor，而通过torch.detach()创建的z则无需进行梯度计算。
5. 总结
本文详细介绍了torch.detach()函数，包括其定义、使用场景以及与.requires_grad_的区别。总的来说，torch.detach()在深度学习中有着广泛的应用，特别是在需要保存不参与梯度计算的Tensor时。
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ ᅟᅠ
