Pytorch:初始化

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]])
# =============================================================================