pytorch中的weight-initilzation用法

pytorch中的权值初始化

官方论坛对weight-initilzation的讨论

torch.nn.Module.apply(fn)

torch.nn.Module.apply(fn)
# 递归的调用weights_init函数,遍历nn.Module的submodule作为参数
# 常用来对模型的参数进行初始化
# fn是对参数进行初始化的函数的句柄,fn以nn.Module或者自己定义的nn.Module的子类作为参数
# fn (Module -> None) – function to be applied to each submodule
# Returns: self
#Return type: Module

例子:

def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
 m.weight.data.normal_(0.0, 0.02)
 # m.weight.data是卷积核参数, m.bias.data是偏置项参数
elif classname.find('BatchNorm') != -1:
 m.weight.data.normal_(1.0, 0.02)
 m.bias.data.fill_(0)

netG = _netG(ngpu)# 生成模型实例
netG.apply(weights_init)# 递归的调用weights_init函数,遍历netG的submodule作为参数

#-*-coding:utf-8-*-
import torch
from torch.autograd import Variable

#对模型参数进行初始化
# 官方论坛链接：https://discuss.pytorch.org/t/weight-initilzation/157/3

# 方法一
# 单独定义一个weights_init函数,输入参数是m(torch.nn.module或者自己定义的继承nn.module的子类)
#然后使用net.apply()进行参数初始化
#m.__class__.__name__获得nn.module的名字
#https://github.com/pytorch/examples/blob/master/dcgan/main.py#L90-L96
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
 m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
 m.weight.data.normal_(1.0, 0.02)
 m.bias.data.fill_(0)

netG = _netG(ngpu)# 生成模型实例
netG.apply(weights_init)# 递归的调用weights_init函数,遍历netG的submodule作为参数

# function to be applied to each submodule

# 方法二
# 1. 使用net.modules()遍历模型中的网络层的类型 2. 对其中的m层的weigth.data(tensor)部分进行初始化操作
# Another initialization example from PyTorch Vision resnet implementation.
# https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py#L112-L118
class ResNet(nn.Module):
def __init__(self, block, layers, num_classes=1000):
 self.inplanes = 64
 super(ResNet, self).__init__()
 self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
       bias=False)
 self.bn1 = nn.BatchNorm2d(64)
 self.relu = nn.ReLU(inplace=True)
 self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
 self.layer1 = self._make_layer(block, 64, layers[0])
 self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
 self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
 self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
 self.avgpool = nn.AvgPool2d(7, stride=1)
 self.fc = nn.Linear(512 * block.expansion, num_classes)
 #权值参数初始化
 for m in self.modules():
  if isinstance(m, nn.Conv2d):
   n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
   m.weight.data.normal_(0, math.sqrt(2. / n))
  elif isinstance(m, nn.BatchNorm2d):
   m.weight.data.fill_(1)
   m.bias.data.zero_()

# 方法三
# 自己知道网络中参数的顺序和类型, 然后将参数依次读取出来,调用torch.nn.init中的方法进行初始化
net = AlexNet(2)
params = list(net.parameters()) #params依次为Conv2d参数和Bias参数
# 或者
conv1Params = list(net.conv1.parameters())
# 其中,conv1Params[0]表示卷积核参数, conv1Params[1]表示bias项参数
# 然后使用torch.nn.init中函数进行初始化
torch.nn.init.normal(tensor, mean=0, std=1)
torch.nn.init.constant(tensor, 0)

# net.modules()迭代的返回: AlexNet,Sequential,Conv2d,ReLU,MaxPool2d,LRN,AvgPool3d....,Conv2d,...,Conv2d,...,Linear,
# 这里,只有Conv2d和Linear才有参数
# net.children()只返回实际存在的子模块: Sequential,Sequential,Sequential,Sequential,Sequential,Sequential,Sequential,Linear

# 附AlexNet的定义
class AlexNet(nn.Module):
def __init__(self, num_classes = 2): # 默认为两类，猫和狗
#   super().__init__() # python3
 super(AlexNet, self).__init__()
 # 开始构建AlexNet网络模型，5层卷积，3层全连接层
 # 5层卷积层
 self.conv1 = nn.Sequential(
  nn.Conv2d(in_channels=3, out_channels=96, kernel_size=11, stride=4),
  nn.ReLU(inplace=True),
  nn.MaxPool2d(kernel_size=3, stride=2),
  LRN(local_size=5, bias=1, alpha=1e-4, beta=0.75, ACROSS_CHANNELS=True)
 )
 self.conv2 = nn.Sequential(
  nn.Conv2d(in_channels=96, out_channels=256, kernel_size=5, groups=2, padding=2),
  nn.ReLU(inplace=True),
  nn.MaxPool2d(kernel_size=3, stride=2),
  LRN(local_size=5, bias=1, alpha=1e-4, beta=0.75, ACROSS_CHANNELS=True)
 )
 self.conv3 = nn.Sequential(
  nn.Conv2d(in_channels=256, out_channels=384, kernel_size=3, padding=1),
  nn.ReLU(inplace=True)
 )
 self.conv4 = nn.Sequential(
  nn.Conv2d(in_channels=384, out_channels=384, kernel_size=3, padding=1),
  nn.ReLU(inplace=True)
 )
 self.conv5 = nn.Sequential(
  nn.Conv2d(in_channels=384, out_channels=256, kernel_size=3, padding=1),
  nn.ReLU(inplace=True),
  nn.MaxPool2d(kernel_size=3, stride=2)
 )
 # 3层全连接层
 # 前向计算的时候，最开始输入需要进行view操作，将3D的tensor变为1D
 self.fc6 = nn.Sequential(
  nn.Linear(in_features=6*6*256, out_features=4096),
  nn.ReLU(inplace=True),
  nn.Dropout()
 )
 self.fc7 = nn.Sequential(
  nn.Linear(in_features=4096, out_features=4096),
  nn.ReLU(inplace=True),
  nn.Dropout()
 )
 self.fc8 = nn.Linear(in_features=4096, out_features=num_classes)

def forward(self, x):
 x = self.conv5(self.conv4(self.conv3(self.conv2(self.conv1(x)))))
 x = x.view(-1, 6*6*256)
 x = self.fc8(self.fc7(self.fc6(x)))
 return x

补充知识：pytorch Load部分weights

我们从网上down下来的模型与我们的模型可能就存在一个层的差异，此时我们就需要重新训练所有的参数是不合理的。

因此我们可以加载相同的参数，而忽略不同的参数，代码如下：

 pretrained_dict = torch.load(“model.pth”)
 model_dict = et.state_dict()
 pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
 model_dict.update(pretrained_dict)
 net.load_state_dict(model_dict)

来源：https://blog.csdn.net/tsq292978891/article/details/79382306