Pytorch -- use and modification of existing network model

         With VGG Take the Internet for example , stay Docs in , Select image related torchvision, Left side torchvision.models Are some ready-made network models , On the right is a classification of network models , The first is the model related to classification .

        vgg The most common is vgg16 and vgg19：

         Parameters pretrained If True, Then the parameters in the model use the existing parameters that are already in ImageNet Data set training is completed , if False, Then the parameter has not been trained ,process if True, A download progress bar will be displayed .

          need scipy The package can be installed ,root Is the path ,split Whether to choose training set or test set ,transform Yes PIL Form transformation ,target_transform Transform your goals ,loader Use... When loading datasets .

train_data = torchvision.datasets.ImageNet("./data_ImageNet", split="train", download=True, transform=torchvision.transforms.ToTensor())

  If you use this method to download , Will report a mistake ：

RuntimeError: The dataset is no longer publicly accessible. You need to download the archives externally and place them in the root directory.

         Data sets are no longer publicly accessible , To download the data set and put it in the directory . Just search the Internet , The size of the training set is 100 Multiple G.

         We load the model directly ：

import torchvision

vgg16_False = torchvision.models.vgg16(pretrained=False)#progress The default is True
vgg16_True = torchvision.models.vgg16(pretrained=True)

View the parameters of the two models ：

vgg16_False:

vgg16_True:

Set to False when , The parameter is the initial default value , by True when , The parameters have been trained , You can get good results . Print out the trained network model ：

VGG(
  (features): Sequential(
    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU(inplace=True)
    (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (6): ReLU(inplace=True)
    (7):...

Look at the last layer ：

Linear(in_features=4096, out_features=1000, bias=True)

The output is 1000 individual , The network is a classification model , We know that in the end 1000 Classes . As mentioned in the previous blog CIFAR10 Data sets divide images into 10 Classes , Can be vgg16 Last linear layer 1000 Change it to 10, Or add a new linear layer , Input 1000, Output 10.

vgg16_True.add_module('add_linear', nn.Linear(1000, 100))

see add_module Source code ：

    def add_module(self, name: str, module: Optional['Module']) -> None:
        r"""Adds a child module to the current module.

        The module can be accessed as an attribute using the given name.

        Args:
            name (string): name of the child module. The child module can be
                accessed from this module using the given name
            module (Module): child module to be added to the module.
        """

name It's the name of the new layer ,module Is the newly added layer . New network is ：

...

  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
  (add_linear): Linear(in_features=1000, out_features=100, bias=True)

If you want to add to classifier in , It is amended as follows ：

vgg16_True.classifier.add_module('add_linear', nn.Linear(1000, 100))

  If you want to put the last one Linear4096 Input ,1000 The output is modified to 4096 Input 10 Output ：

vgg16_True.classifier[6] = nn.Linear(4096, 10)

（ Because the last floor is classifier in , And the serial number is 6）