【Pytorch学习】|torch.nn.init常用函数总结深度学习|python|pytorch

torch.nn.init中常用的函数：

torch.nn.init.uniform_

torch.nn.init.normal_

torch.nn.init.constant_

torch.nn.init.ones_

torch.nn.init.zeros_

torch.nn.init.eye_

torch.nn.init.sparse_

torch.nn.init torch.nn.init.uniform_
torch.nn.init.uniform_(tensor, a=0.0, b=1.0)
按均匀分布 U ( a , b ) U(a,b) U(a,b)对tensor随机赋值。

import torch import torch.nn as nnclass MyModel(nn.Module): def __init__(self, **kwargs): super(MyModel, self).__init__()self.linear = nn.Linear(in_features=3, out_features=1) # 此时输出为随机赋值的参数值 print(self.linear.weight) print(self.linear.bias) print() # 使用nn.Parameter对参数赋值 self.linear.weight = nn.Parameter(torch.ones(1, 3, dtype=torch.float32)) self.linear.bias = nn.Parameter(torch.ones(1, dtype=torch.float32)) print(self.linear.weight) print(self.linear.bias) print() # 使用nn.init.uniform_对参数赋值 nn.init.uniform_(self.linear.weight) nn.init.uniform_(self.linear.bias) print(self.linear.weight) print(self.linear.bias)def forward(self, x): passn = MyModel() print() for name, param in n.named_parameters(): # 显示注册的参数 print(name) print(param)""" Parameter containing: tensor([[ 0.2034, -0.0841,0.5076]], requires_grad=True) Parameter containing: tensor([-0.1976], requires_grad=True)Parameter containing: tensor([[1., 1., 1.]], requires_grad=True) Parameter containing: tensor([1.], requires_grad=True)Parameter containing: tensor([[0.3873, 0.0731, 0.8832]], requires_grad=True) Parameter containing: tensor([0.1925], requires_grad=True)linear.weight Parameter containing: tensor([[0.3873, 0.0731, 0.8832]], requires_grad=True) linear.bias Parameter containing: tensor([0.1925], requires_grad=True) """

【【Pytorch学习】|torch.nn.init常用函数总结】查看源码可以发现这么一句：with torch.no_grad():，这样一来我们就不需要自己写这句话了。（下面的函数同）
torch.nn.init.normal_
torch.nn.init.normal_(tensor, mean=0.0, std=1.0)
按正态分布 N ( μ 或 m e a n , σ 或 s t d ) N(\mu或mean,\sigma或std) N(μ或mean,σ或std)对tensor随机赋值。
torch.nn.init.constant_
torch.nn.init.constant_(tensor, val)
使用常数val对tensor赋值。

import torch.nn as nnclass MyModel(nn.Module): def __init__(self, **kwargs): super(MyModel, self).__init__()self.linear = nn.Linear(in_features=3, out_features=1) # 使用nn.init.constant_对参数赋值 nn.init.constant_(self.linear.weight, 0.2) nn.init.constant_(self.linear.bias, 0.2) print(self.linear.weight) print(self.linear.bias)def forward(self, x): passn = MyModel() """ Parameter containing: tensor([[0.2000, 0.2000, 0.2000]], requires_grad=True) Parameter containing: tensor([0.2000], requires_grad=True) """

torch.nn.init.ones_
torch.nn.init.ones_(tensor)
使用常数1对tensor赋值。
torch.nn.init.zeros_
torch.nn.init.zeros_(tensor)
使用常数0对tensor赋值。
torch.nn.init.eye_
torch.nn.init.eye_(tensor)
使用常数1对tensor的主对角线进行赋值，其余位置为常数0。
torch.nn.init.sparse_
torch.nn.init.sparse_(tensor, sparsity, std=0.01)
生成稀疏矩阵。
sparsity ：每列中需要被设置成零的元素比例。
std ：用于生成的正态分布的标准差。

import torch import torch.nn as nnw = torch.empty(8, 5, dtype=torch.float32) print(nn.init.sparse_(w, sparsity=0.5)) """ tensor([[-0.0058, -0.0074, -0.0078, -0.0055,0.0000], [ 0.0000, -0.0023,0.0000,0.0014,0.0046], [ 0.0000,0.0000,0.0000,0.0003, -0.0004], [-0.0029,0.0000,0.0000,0.0000,0.0000], [ 0.0000, -0.0127,0.0000, -0.0078, -0.0174], [-0.0017,0.0000,0.0049,0.0000,0.0000], [ 0.0000,0.0000, -0.0067,0.0000,0.0072], [ 0.0091,0.0027, -0.0008,0.0000,0.0000]]) """