【torch.nn.init】初始化参数方法解读

 可参考：

一. torch.nn.init.constant_(tensor, val)  

1. 作用：

        常数分布： 用值val填充向量。

2. 参数：

• tensor – an n-dimensional torch.Tensor
• val – the value to fill the tensor with

3. 实例：

import torch
form torch from nn
w = torch.empty(3, 5)
print(w)
print(nn.init.constant_(w, 0.3))

-------------------------------------
tensor([[6.4069e+02, 2.74e+20, 1.5444e+25, 1.6217e-19, 7.0062e+22],
        [1.6795e+08, 4.7423e+30, 4.7393e+30, 9.5461e-01, 4.4377e+27],
        [1.7975e+19, 4.e+27, 7.9463e+08, 3.2604e-12, 2.6209e+20]])
tensor([[0.3000, 0.3000, 0.3000, 0.3000, 0.3000],
        [0.3000, 0.3000, 0.3000, 0.3000, 0.3000],
        [0.3000, 0.3000, 0.3000, 0.3000, 0.3000]])

二. torch.nn.init.normal_(tensor, mean=0, std=1)

1. 作用：

        正态分布：从给定均值和标准差的正态分布N(mean, std)中生成值，填充输入的张量或变量。

2. 参数：

• tensor – n维的torch.Tensor
• mean – 正态分布的均值
• std – 正态分布的标准差

3. 实例：

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(torch.nn.init.normal_(w))

----------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[-1.1406, -0.1720, -1.4460,  0.5305, -0.0854],
        [ 0.92,  0.3495, -0.8262, -1.41, -0.26],
        [ 0.7404,  0.7124, -0.3902,  0.0625,  0.6256]])

三. torch.nn.init.uniform_(tensor, a=0.0, b=1.0)

1.作用：

       均匀分布： 从均匀分布N(a,b)中生成值，填充输入的张量或变量。

2. 参数：

• tensor – n 维的torch.Tensor
• a – 均匀分布的下界
• b – 均匀分布的上界

3. 实例：

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.uniform_(w))

-----------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[0.6653, 0.9605, 0.2208, 0.0140, 0.9672],
        [0.4201, 0.5819, 0.8383, 0.4334, 0.0673],
        [0.1246, 0.4066, 0.3413, 0.1231, 0.0463]])

四. torch.nn.init.ones_(tensor)

1.作用：

  全1分布：用标量值 1 填充输入张量。

2. 参数：

• tensor – n 维的torch.Tensor

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.ones_(w))

---------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.]])

五. torch.nn.init.zeros_(tensor)

1.作用：

        全0分布：用全0填充张量。

2. 参数： 

• tensor – n 维的torch.Tensor

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.zeros_(w))

-------------------------------------------------

tensor([[-4.2990e-27,  4.5701e-41, -4.2990e-27,  4.5701e-41,         nan],
        [ 4.5699e-41,  7.6194e+31,  1.55e+28,  4.7984e+30,  6.2121e+22],
        [ 1.8370e+25,  1.4603e-19,  6.4069e+02,  2.74e+20,  1.5444e+25]])
tensor([[0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0.]])

六. torch.nn.init.eye_(tensor)

1.作用：

  对角分布：用来填充2维输入张量或变量。

2. 参数：

• tensor – 2维的torch.Tensor 或 autograd.Variable

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.eye_(w))

-------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[1., 0., 0., 0., 0.],
        [0., 1., 0., 0., 0.],
        [0., 0., 1., 0., 0.]])

七. torch.nn.init.dirac_(tensor, groups=1)

1.作用：

  dirac分布：用Dirac δ函数来填充{3, 4, 5}维输入张量或变量。在尽可能多的保存输入通道特性。

2. 参数：

• tensor – {3, 4, 5}维的torch.Tensor 或 autograd.Variable

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 16, 5, 5)
print(w.shape)
print(nn.init.dirac_(w).shape)
z = torch.empty(3, 24, 5, 5)
print(z.shape)
print(nn.init.dirac_(z, 3).shape)

---------------------------------------------

torch.Size([3, 16, 5, 5])
torch.Size([3, 16, 5, 5])
torch.Size([3, 24, 5, 5])
torch.Size([3, 24, 5, 5])

八. torch.nn.init.xavier_uniform_(tensor, gain=1.0)

1. 作用： 

        xavier_uniform分布：用一个均匀分布生成值，填充输入的张量或变量。

2. 参数： 

• tensor – n维的torch.Tensor
• gain – 可选的缩放因子

 3.实例：

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain('relu')))

----------------------------------------------------------

tensor([[6.4069e+02, 2.74e+20, 1.5444e+25, 1.6217e-19, 7.0062e+22],
        [1.6795e+08, 4.7423e+30, 4.7393e+30, 9.5461e-01, 4.4377e+27],
        [1.7975e+19, 4.e+27, 7.9463e+08, 3.2604e-12, 2.6209e+20]])
tensor([[-0.9562, -0.6834,  0.7449, -0.2484, -0.7638],
        [-1.0150, -0.2982, -0.2133, -1.1132, -1.0273],
        [ 0.5228,  0.9122, -0.5077, -0.2911,  0.1625]])

 九. torch.nn.init.xavier_normal_(tensor, gain=1.0)

1. 作用： 

         xavier_normal 分布：用一个正态分布生成值，填充输入的张量或变量。

 2. 参数：  

• tensor – n维的torch.Tensor
• gain – 可选的缩放因子

 3. 实例：

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.xavier_normal_(w))

--------------------------------------------

tensor([[4.7984e+30, 6.2121e+22, 1.8370e+25, 1.4603e-19, 6.4069e+02],
        [2.74e+20, 1.5444e+25, 1.6217e-19, 7.0062e+22, 1.6795e+08],
        [4.7423e+30, 4.7393e+30, 9.5461e-01, 4.4377e+27, 1.7975e+19]])
tensor([[ 0.3654,  0.4767,  0.1407, -0.4990,  0.2799],
        [ 0.0545,  0.5941, -0.3611,  0.5469,  0.0781],
        [-0.0393,  0.1817, -0.0407, -0.2593, -0.2736]])

 十. torch.nn.init.kaiming_uniform_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu')

1. 作用：

        kaiming_uniform 分布：用一个均匀分布生成值，填充输入的张量或变量。 

2. 参数： 

• tensor – n维的torch.Tensor或autograd.Variable；
• a – 这层之后使用的rectifier的斜率系数（ReLU的默认值为0）；
• mode – 可以为“fan_in”（默认）或 “fan_out”；
• “fan_in” – 保留前向传播时权值方差的量级；
• “fan_out” – 保留反向传播时的量级；
• nonlinearity=‘leaky_relu’ – 非线性函数 建议“relu”或“leaky_relu”（默认值）使用。

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.kaiming_uniform_(w, mode='fan_in', nonlinearity='relu'))

-------------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[ 0.6771, -0.7587,  0.6915, -0.7163,  0.0840],
        [-1.0694, -0.4790, -0.4019, -0.8439,  0.5794],
        [-0.9363, -0.0655, -0.0506, -0.1419,  0.5395]])

十一. torch.nn.init.kaiming_normal_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu')

1. 作用： 

        kaiming_normal 分布：用一个正态分布生成值，填充输入的张量或变量。 

2. 参数： 

• tensor – n维的torch.Tensor或autograd.Variable；
• a – 这层之后使用的rectifier的斜率系数（ReLU的默认值为0）；
• mode – 可以为“fan_in”（默认）或 “fan_out”fan_in保留前向传播时权值方差的量级fan_out保留反向传播时的量级。

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.kaiming_normal_(w, mode='fan_out', nonlinearity='relu'))

-------------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[-0.2421,  1.3102, -0.0506,  0.5099, -0.1017],
        [-1.2707, -0.9636, -0.4539,  1.1167,  0.6717],
        [ 0.18,  0.6261, -1.1114, -0.4440,  0.5798]])

十二. torch.nn.init.orthogonal_(tensor, gain=1) 

 1. 作用：

        正交矩阵：用一个(半)正交矩阵填充输入张量。 

2. 参数： 

• tensor– 一个n维的tensor,其中 n≥2
• gain– 可选比例系数

3. 实例： 

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.orthogonal_(w))

------------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[-0.2146, -0.87, -0.3447, -0.1060,  0.2363],
        [-0.1957,  0.2711,  0.0974, -0.38,  0.6813],
        [ 0.6258, -0.3716,  0.6203, -0.2903, -0.0353]]

十二. torch.nn.init.sparse_(tensor, sparsity, std=0.01) 

1. 作用： 

        稀疏矩阵：将2D输入张量填充为稀疏矩阵，其中非零元素将从正态分布N ( 0 , 0.01 ) N(0,0.01)N(0,0.01)中提取。

 2. 参数：

• tensor– 一个n维的torch.tensor张量
• sparsity– 每一列中元素的比例设置为零
• std– 用于产生非零值的正态分布的标准差

 3. 实例：

import torch
from torch import nn
w = torch.empty(3, 5)
print(w)
print(nn.init.sparse_(w, sparsity=0.1))

------------------------------------------

tensor([[9.5461e-01, 4.4377e+27, 1.7975e+19, 4.e+27, 7.9463e+08],
        [3.2604e-12, 2.6209e+20, 4.11e+12, 1.9434e-19, 3.0881e+29],
        [6.3828e+28, 1.4603e-19, 7.7179e+28, 7.7591e+26, 3.0357e+32]])
tensor([[ 0.0112,  0.0000, -0.0055,  0.0000,  0.0000],
        [ 0.0026, -0.0009,  0.0000, -0.0044, -0.0012],
        [ 0.0000,  0.0176,  0.0022, -0.0037, -0.0035]])