标签：初始化 tensor nn torch init Pytorch 0.3000
4.4 初始化策略
在深度学习中参数的初始化十分重要，良好的初始化能让模型更快收敛，并达到更高水平，而糟糕的初始化则可能使得模型迅速瘫痪。PyTorch中nn.Module的模块参数都采取了较为合理的初始化策略，因此一般不用我们考虑，当然我们也可以用自定义初始化去代替系统的默认初始化。而当我们在使用Parameter时，自定义初始化则尤为重要，因t.Tensor()返回的是内存中的随机数，很可能会有极大值，这在实际训练网络中会造成溢出或者梯度消失。PyTorch中nn.init模块就是专门为初始化而设计，如果某种初始化策略nn.init不提供，用户也可以自己直接初始化。
In [55]:
# 利用nn.init初始化
from torch.nn import init
linear = nn.Linear(3, 4)
t.manual_seed(1)
# 等价于 linear.weight.data.normal_(0, std)
init.xavier_normal_(linear.weight)
Out[55]:
Parameter containing:
tensor([[ 0.3535, 0.1427, 0.0330],
[ 0.3321, -0.2416, -0.0888],
[-0.8140, 0.2040, -0.5493],
[-0.3010, -0.4769, -0.0311]])In [56]:
# 直接初始化
import math
t.manual_seed(1)
# xavier初始化的计算公式
std = math.sqrt(2)/math.sqrt(7.)
linear.weight.data.normal_(0,std)
Out[56]:
tensor([[ 0.3535, 0.1427, 0.0330],
[ 0.3321, -0.2416, -0.0888],
[-0.8140, 0.2040, -0.5493],
[-0.3010, -0.4769, -0.0311]])In [57]:
# 对模型的所有参数进行初始化
for name, params in net.named_parameters():
if name.find('linear') != -1:
# init linear
params[0] # weight
params[1] # bias
elif name.find('conv') != -1:
pass
elif name.find('norm') != -1:
pass
初始化的具体例子
# -*- coding: utf-8 -*-
"""
Created on 2019
@author: fancp
"""
import torch
import torch.nn as nn
w = torch.empty(3,5)
#1.均匀分布 - u(a,b)
#torch.nn.init.uniform_(tensor, a=0.0, b=1.0)
print(nn.init.uniform_(w))
# =============================================================================
# tensor([[0.9160, 0.1832, 0.5278, 0.5480, 0.6754],
# [0.9509, 0.8325, 0.9149, 0.8192, 0.9950],
# [0.4847, 0.4148, 0.8161, 0.0948, 0.3787]])
# =============================================================================
#2.正态分布 - N(mean, std)
#torch.nn.init.normal_(tensor, mean=0.0, std=1.0)
print(nn.init.normal_(w))
# =============================================================================
# tensor([[ 0.4388, 0.3083, -0.6803, -1.1476, -0.6084],
# [ 0.5148, -0.2876, -1.2222, 0.6990, -0.1595],
# [-2.0834, -1.6288, 0.5057, -0.5754, 0.3052]])
# =============================================================================
#3.常数 - 固定值 val
#torch.nn.init.constant_(tensor, val)
print(nn.init.constant_(w, 0.3))
# =============================================================================
# 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]])
# =============================================================================
#4.全1分布
#torch.nn.init.ones_(tensor)
print(nn.init.ones_(w))
# =============================================================================
# tensor([[1., 1., 1., 1., 1.],
# [1., 1., 1., 1., 1.],
# [1., 1., 1., 1., 1.]])
# =============================================================================
#5.全0分布
#torch.nn.init.zeros_(tensor)
print(nn.init.zeros_(w))
# =============================================================================
# tensor([[0., 0., 0., 0., 0.],
# [0., 0., 0., 0., 0.],
# [0., 0., 0., 0., 0.]])
# =============================================================================
#6.对角线为 1，其它为 0
#torch.nn.init.eye_(tensor)
print(nn.init.eye_(w))
# =============================================================================
# tensor([[1., 0., 0., 0., 0.],
# [0., 1., 0., 0., 0.],
# [0., 0., 1., 0., 0.]])
# =============================================================================
#7.xavier_uniform 初始化
#torch.nn.init.xavier_uniform_(tensor, gain=1.0)
#From - Understanding the difficulty of training deep feedforward neural networks - Bengio 2010
print(nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain('relu')))
# =============================================================================
# tensor([[-0.1270, 0.3963, 0.9531, -0.2949, 0.8294],
# [-0.9759, -0.6335, 0.9299, -1.0988, -0.1496],
# [-0.7224, 0.2181, -1.1219, 0.8629, -0.8825]])
# =============================================================================
#8.xavier_normal 初始化
#torch.nn.init.xavier_normal_(tensor, gain=1.0)
print(nn.init.xavier_normal_(w))
# =============================================================================
# tensor([[ 1.0463, 0.1275, -0.3752, 0.1858, 1.1008],
# [-0.5560, 0.2837, 0.1000, -0.5835, 0.7886],
# [-0.2417, 0.1763, -0.7495, 0.4677, -0.1185]])
# =============================================================================
#9.kaiming_uniform 初始化
#torch.nn.init.kaiming_uniform_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu')
#From - Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification - HeKaiming 2015
print(nn.init.kaiming_uniform_(w, mode='fan_in', nonlinearity='relu'))
# =============================================================================
# tensor([[-0.7712, 0.9344, 0.8304, 0.2367, 0.0478],
# [-0.6139, -0.3916, -0.0835, 0.5975, 0.1717],
# [ 0.3197, -0.9825, -0.5380, -1.0033, -0.3701]])
# =============================================================================
#10.kaiming_normal 初始化
#torch.nn.init.kaiming_normal_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu')
print(nn.init.kaiming_normal_(w, mode='fan_out', nonlinearity='relu'))
# =============================================================================
# tensor([[-0.0210, 0.5532, -0.8647, 0.9813, 0.0466],
# [ 0.7713, -1.0418, 0.7264, 0.5547, 0.7403],
# [-0.8471, -1.7371, 1.3333, 0.0395, 1.0787]])
# =============================================================================
#11.正交矩阵 - (semi)orthogonal matrix
#torch.nn.init.orthogonal_(tensor, gain=1)
#From - Exact solutions to the nonlinear dynamics of learning in deep linear neural networks - Saxe 2013
print(nn.init.orthogonal_(w))
# =============================================================================
# tensor([[-0.0346, -0.7607, -0.0428, 0.4771, 0.4366],
# [-0.0412, -0.0836, 0.9847, 0.0703, -0.1293],
# [-0.6639, 0.4551, 0.0731, 0.1674, 0.5646]])
# =============================================================================
#12.稀疏矩阵 - sparse matrix
#torch.nn.init.sparse_(tensor, sparsity, std=0.01)
#From - Deep learning via Hessian-free optimization - Martens 2010
print(nn.init.sparse_(w, sparsity=0.1))
# =============================================================================
# tensor([[ 0.0000, 0.0000, -0.0077, 0.0000, -0.0046],
# [ 0.0152, 0.0030, 0.0000, -0.0029, 0.0005],
# [ 0.0199, 0.0132, -0.0088, 0.0060, 0.0000]])
# =============================================================================
标签：初始化,tensor,nn,torch,init,Pytorch,0.3000 来源： https://blog.csdn.net/qimo601/article/details/112720177
本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享； 2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关； 3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关； 4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除； 5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。