当前位置：首页 > article >正文

警告 torch.nn.utils.weight_norm is deprecate 的参考解决方法

article 2025/1/20 16:31:37

文章目录

写在前面
一、问题描述
二、解决方法
参考链接

写在前面

自己的测试环境：
Ubuntu20.04

一、问题描述

运行 pytorch 程序，如下如下警告

/home/wong/ProgramFiles/anaconda3/envs/pytorch_env/lib/python3.8/site-packages/torch/nn/utils/weight_norm.py:30: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
  warnings.warn("torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.")

在这里插入图片描述

二、解决方法

根据提示可以发现，torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm. 也就是 torch.nn.utils.weight_norm 已经被弃用，建议使用 torch.nn.utils.parametrizations.weight_norm.
因此需要改为 torch.nn.utils.parametrizations.weight_norm , 并且，根据pytorch官网中的描述，还需要把程序中的 weight_g 修改为parametrizations.weight.original0， weight_v 修改为parametrizations.weight.original1 .

比如，原程序为

theta = torch.nn.utils.weight_norm(nn.Linear(d_in, num_projections, bias=False), dim=0)
if num_projections <= d_in:
    torch.nn.init.eye_(theta.weight_v)
else:
    torch.nn.init.normal_(theta.weight_v)
theta.weight_g.data = torch.ones_like(theta.weight_g.data, requires_grad=False)
theta.weight_g.requires_grad = False

需要修改为：

theta = torch.nn.utils.parametrizations.weight_norm(nn.Linear(d_in, num_projections, bias=False), dim=0)
if num_projections <= d_in:
    torch.nn.init.eye_(theta.parametrizations.weight.original1)
else:
    torch.nn.init.normal_(theta.parametrizations.weight.original1)
theta.parametrizations.weight.original0.data = torch.ones_like(theta.parametrizations.weight.original0.data, requires_grad=False)
theta.parametrizations.weight.original0.requires_grad = False