快捷導(dǎo)航

Pytorch - TORCH.NN.INIT 參數(shù)初始化的操作

更新時(shí)間：2021年02月27日 12:11:28 作者：Peanut_范

這篇文章主要介紹了Pytorch - TORCH.NN.INIT 參數(shù)初始化的操作，具有很好的參考價(jià)值，希望對(duì)大家有所幫助。一起跟隨小編過來看看吧

路徑：

https://pytorch.org/docs/master/nn.init.html#nn-init-doc

初始化函數(shù)：torch.nn.init

# -*- coding: utf-8 -*-
"""
Created on 2019
@author: fancp
"""
import torch 
import torch.nn as nn
w = torch.empty(3,5)
#1.均勻分布 - 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.正態(tài)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.常數(shù) - 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.對(duì)角線為 1，其它為 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.正交矩陣 - (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.稀疏矩陣 - 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]])
# =============================================================================

補(bǔ)充：【pytorch參數(shù)初始化】 pytorch默認(rèn)參數(shù)初始化以及自定義參數(shù)初始化

本文用兩個(gè)問題來引入

1.pytorch自定義網(wǎng)絡(luò)結(jié)構(gòu)不進(jìn)行參數(shù)初始化會(huì)怎樣，參數(shù)值是隨機(jī)的嗎？

2.如何自定義參數(shù)初始化？

先回答第一個(gè)問題

在pytorch中，有自己默認(rèn)初始化參數(shù)方式，所以在你定義好網(wǎng)絡(luò)結(jié)構(gòu)以后，不進(jìn)行參數(shù)初始化也是可以的。

1.Conv2d繼承自_ConvNd，在_ConvNd中，可以看到默認(rèn)參數(shù)就是進(jìn)行初始化的，如下圖所示

2.torch.nn.BatchNorm2d也一樣有默認(rèn)初始化的方式

3.torch.nn.Linear也如此

現(xiàn)在來回答第二個(gè)問題。

pytorch中對(duì)神經(jīng)網(wǎng)絡(luò)模型中的參數(shù)進(jìn)行初始化方法如下：

from torch.nn import init
#define the initial function to init the layer's parameters for the network
def weigth_init(m):
  if isinstance(m, nn.Conv2d):
    init.xavier_uniform_(m.weight.data)
    init.constant_(m.bias.data,0.1)
  elif isinstance(m, nn.BatchNorm2d):
    m.weight.data.fill_(1)
    m.bias.data.zero_()
  elif isinstance(m, nn.Linear):
    m.weight.data.normal_(0,0.01)
    m.bias.data.zero_()

首先定義了一個(gè)初始化函數(shù)，接著進(jìn)行調(diào)用就ok了，不過要先把網(wǎng)絡(luò)模型實(shí)例化：

 #Define Network
  model = Net(args.input_channel,args.output_channel)
  model.apply(weigth_init)

此上就完成了對(duì)模型中訓(xùn)練參數(shù)的初始化。

在知乎上也有看到一個(gè)類似的版本，也相應(yīng)的貼上來作為參考了：

def initNetParams(net):
  '''Init net parameters.'''
  for m in net.modules():
    if isinstance(m, nn.Conv2d):
      init.xavier_uniform(m.weight)
      if m.bias:
        init.constant(m.bias, 0)
    elif isinstance(m, nn.BatchNorm2d):
      init.constant(m.weight, 1)
      init.constant(m.bias, 0)
    elif isinstance(m, nn.Linear):
      init.normal(m.weight, std=1e-3)
      if m.bias:
        init.constant(m.bias, 0) 
initNetParams(net)

再說一下關(guān)于模型的保存及加載

1.保存有兩種方式，第一種是保存模型的整個(gè)結(jié)構(gòu)信息和參數(shù)，第二種是只保存模型的參數(shù)

 #保存整個(gè)網(wǎng)絡(luò)模型及參數(shù)
 torch.save(net, 'net.pkl') 
 
 #僅保存模型參數(shù)
 torch.save(net.state_dict(), 'net_params.pkl')

2.加載對(duì)應(yīng)保存的兩種網(wǎng)絡(luò)

# 保存和加載整個(gè)模型 
torch.save(model_object, 'model.pth') 
model = torch.load('model.pth') 
 
# 僅保存和加載模型參數(shù) 
torch.save(model_object.state_dict(), 'params.pth') 
model_object.load_state_dict(torch.load('params.pth'))

以上為個(gè)人經(jīng)驗(yàn)，希望能給大家一個(gè)參考，也希望大家多多支持腳本之家。如有錯(cuò)誤或未考慮完全的地方，望不吝賜教。

您可能感興趣的文章: