可参考：torch.nn.init – 云+社区 – 腾讯云
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, 2.7489e+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, 4.6894e+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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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],
[ 0.8992, 0.3495, -0.8262, -1.4641, -0.6426],
[ 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.Tensora– 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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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, 1.5564e+28, 4.7984e+30, 6.2121e+22],
[ 1.8370e+25, 1.4603e-19, 6.4069e+02, 2.7489e+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：用单位矩阵来填充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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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。在卷积层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.Tensorgain– 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, 2.7489e+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, 4.6894e+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.Tensorgain– 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],
[2.7489e+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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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],
[ 0.1898, 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≥2gain– 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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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, -0.8764, -0.3447, -0.1060, 0.2363],
[-0.1957, 0.2711, 0.0974, -0.6438, 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, 4.6894e+27, 7.9463e+08],
[3.2604e-12, 2.6209e+20, 4.1641e+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]])文章出处登录后可见！
已经登录？立即刷新
