[pytorch modifies the pre training model: there is little difference between the measured loading pre training model and the random initialization of the model]

---

1. pytorch Pre training model

The training of convolutional neural network is time-consuming , In many cases, it is impossible to train the network every time from the random initialization parameters .pytorch Several commonly used deep learning network pre training models are included in , Such as VGG、ResNet etc. . Often in order to speed up the progress of learning , At the beginning of training, we directly load pre-train Pre trained parameters in the model ,model The loading of is as follows ：

import torchvision.models as models

#resnet
model = models.ResNet(pretrained=True)
model = models.resnet18(pretrained=True)
model = models.resnet34(pretrained=True)
model = models.resnet50(pretrained=True)

#vgg
model = models.VGG(pretrained=True)
model = models.vgg11(pretrained=True)
model = models.vgg16(pretrained=True)
model = models.vgg16_bn(pretrained=True)

2. Modify the number of full connection layer categories

The pre training model is based on resnet50 For example .

model = torchvision.models.resnet50(pretrained=True)
# extract fc Fixed parameters in layer 
fc_features = model.fc.in_features
# Change the category to 10, Redefine the last layer 
model.fc = nn.Linear(fc_features ,10)
print(model.fc)

Or directly transfer in the number of categories ：

self.resnet = torchvision.models.resnet50(pretrained=False,num_classes=10)

3. Modify the convolution of a certain layer

The pre training model is based on resnet50 For example .

model = torchvision.models.resnet50(pretrained=True)
#  Redefine the number of input channels of the first layer convolution 
model.conv1 = nn.Conv2d(4, 64, kernel_size=7, stride=2, padding=3, bias=False)

4. Modify the convolution of certain layers

4.1 Remove the last two layers (fc Layer and the pooling layer )

The pre training model is based on resnet50 For example .
nn.module Of model It contains a called children() Function of , This function can be used to extract model The network structure of each layer , On this basis, it can be modified , The modification method is as follows ( Remove the last two layers )：

resnet_50_s = torchvision.models.resnet50(pretrained=False)
resnet_layer = nn.Sequential(*list(resnet_50_s.children())[:-2])
self.resnet = resnet_layer

After removing the pre training resnet The last two layers of the model (fc Layer and the pooling layer ) after , Add a new upper sampling layer 、 Pool layer and classification layer , The code to build the network is as follows ：

class Net_resnet50_upsample(nn.Module):
     def __init__(self):
         super(Net_resnet50_upsample, self).__init__()
         self.conv = nn.Conv2d(1, 3, kernel_size=1)
         resnet_50_s = torchvision.models.resnet50(pretrained=False)
         resnet_layer = nn.Sequential(*list(resnet_50_s.children())[:-2])
         self.resnet = resnet_layer
         # print(self.resnet)
         
         self.up7to14=nn.UpsamplingNearest2d(scale_factor=2)
         self.avgpool=nn.AvgPool2d(7,stride=2)
         self.fc = nn.Sequential(
             nn.Linear(2048 * 4 * 4, 1024),
             nn.ReLU(inplace=True),
             nn.Linear(1024, 128),
             nn.ReLU(inplace=True),
             nn.Linear(128, 10))
     def forward(self, x):
         x = self.conv(x)
         x = self.resnet(x)
         x=self.up7to14(x)
         x=self.avgpool(x)
         x = x.view(x.size(0), -1)
         x = self.fc(x)
         return x

4.2 Add or remove multiple convolution layers

Sometimes it is necessary to modify the hierarchy in the network , At this time, only the method of parameter coverage can be used , That is to define a similar network first , Then extract the parameters in the pre training to their own network . Here we use resnet Examples of pre training models .

# coding=UTF-8
import torchvision.models as models
import torch
import torch.nn as nn
import math
import torch.utils.model_zoo as model_zoo
 
#Bottleneck It's a class  It defines the use of 1*1 A residual block for dimension reduction and dimension increase with the convolution kernel of , Can be in github resnet pytorch Check out 
class Bottleneck(nn.Module):
    expansion = 4
 
    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride
 
    def forward(self, x):
        residual = x
 
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
 
        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)
 
        out = self.conv3(out)
        out = self.bn3(out)
 
        if self.downsample is not None:
            residual = self.downsample(x)
 
        out += residual
        out = self.relu(out)
 
        return out
 
# Layers that are not modified cannot be named randomly , Otherwise, the pre training weight parameter cannot be passed in 
class CNN(nn.Module):
    def __init__(self, block, layers, num_classes=9):
        self.inplanes = 64
        super(CNN, 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.AdaptiveAvgPool2d(output_size=(1,1))
        #  Add a new anti roll layer 
        self.convtranspose1 = nn.ConvTranspose2d(2048, 2048, kernel_size=3, stride=1, padding=1, output_padding=0,
                                                 groups=1, bias=False, dilation=1)
        #  Add a maximum pooling layer 
        self.maxpool2 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
        #  Remove the original fc layer , Add a new one fclass layer 
        self.fclass = nn.Linear(2048, num_classes)
 
        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_()
 
    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )
 
        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))
 
        return nn.Sequential(*layers)
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
 
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
 
        x = self.avgpool(x)
        #  Newly added layer forward
        x = x.view(x.size(0), -1)
        x = self.convtranspose1(x)
        x = self.maxpool2(x)
        x = x.view(x.size(0), -1)
        x = self.fclass(x)
 
        return x
 
#  load model
resnet50 = models.resnet50(pretrained=False)
print(resnet50)
cnn = CNN(Bottleneck, [3, 4, 6, 3]) #3 4 6 3  respectively layer1 2 3 4  in Bottleneck The number of modules .res101 Then for 3 4 23 3 
#  Read parameters 
pretrained_dict = resnet50.state_dict()
model_dict = cnn.state_dict()
#  take pretrained_dict It doesn't belong to model_dict Key out of 
pretrained_dict = {
    
    k: v for k, v in pretrained_dict.items() if k in model_dict}
#  Update existing model_dict
model_dict.update(pretrained_dict)
#  Load what we really need state_dict
cnn.load_state_dict(model_dict)
# print(resnet50)
print(cnn)

Results contrast ：

The knowledge points of the article match the official knowledge files , You can further learn relevant knowledge