PyTorch张量的关闭自动梯度的三种方式

敲代码的小风

于 2020-12-29 15:04:08 发布

阅读量1w

点赞数 3

分类专栏：零基础学习SSD网络PyTorch实现 Deep-Learning-with-PyTorch 《深度学习之PyTorch实战计算机视觉》

本文链接：https://blog.csdn.net/m0_46653437/article/details/111915713

版权

零基础学习SSD网络PyTorch实现同时被 3 个专栏收录

293 篇文章

订阅专栏

Deep-Learning-with-PyTorch

216 篇文章

订阅专栏

《深度学习之PyTorch实战计算机视觉》

198 篇文章

订阅专栏

参考链接: PyTorch Tutorial 03 - Gradient Calculation With Autograd

三种方式:

# x.requires_grad_(False)
# x.detach()
# with torch.no_grad():

'''
prevent PyTorch from tracking history and calculating gradients
'''

import torch

torch.manual_seed(seed=20200910)
x = torch.randn(3, requires_grad=True)  
print(x)  # tensor([ 0.2824, -0.3715,  0.9088], requires_grad=True)

# x.requires_grad_(False)
# x.detach()
# with torch.no_grad():


# # 方式1
# x.requires_grad_(False)
# print(x)  # tensor([ 0.2824, -0.3715,  0.9088])

# # 方式2
# y = x.detach()
# print(y)  # tensor([ 0.2824, -0.3715,  0.9088])


# # 方式3
# with torch.no_grad():
#     y = x + 2
#     print(y)  # requires gradients         tensor([2.2824, 1.6285, 2.9088])
# z = x + 2
# print(z)  # don't  requires gradients      tensor([2.2824, 1.6285, 2.9088], grad_fn=<AddBackward0>)