Shulou(Shulou.com)06/03 Report--

This article introduces the relevant knowledge of "what is the pytorch network block VGG". Many people will encounter such a dilemma in the operation of actual cases, so let the editor lead you to learn how to deal with these situations. I hope you can read it carefully and be able to achieve something!

Catalogue

VGG block

VGG network

Training model

Similar to the process of engineers from placing transistors to logic elements and then to logic blocks in chip design, the design of neural network structure has gradually become more abstract. The researchers began to think about the problem from the perspective of a single neuron, developed to the whole level, and now turned to modules, repeating the patterns of each layer.

The idea of using blocks first appeared in the VGG network of the Visual Geometry Group (visualgeometry Group) (VGG) at Oxford University. By using loops and subroutines, these repetitive structures can be easily implemented in the code of any modern deep learning framework.

VGG block

The basic components of the classical convolution neural network are the following sequence:

1. Convolution layer with filling to maintain resolution

two。 Nonlinear activation function, such as ReLU

3. An aggregation layer, such as the maximum aggregation layer.

A VGG block is similar, consisting of a series of convolution layers, followed by a maximum convergence layer for spatial downsampling.

In the original VGG paper, the authors used a convolution layer with a 3 × 3 convolution core filled with 1 (maintaining height and width) and a maximum convergence layer with a 2 × 2 pooling window with a stride of 2 (the resolution after each block was halved).

In the following code, we define a function called vgg_block to implement the VGG block.

This function has three parameters, corresponding to the number of convolution layers num_convs, the number of input channels in_channels and the number of output channels out_channels.

Import torchfrom torch import nnfrom d21 import torch as d2ldef vgg_block (num_convs, in_channels, out_channels): layers = [] for _ in range (num_convs): layers.append (nn.Conv2d (in_channels, out_channels, kernel_size= 3, padding=1)) leyers.append (nn.ReLU ()) in_channels = out_channels layers.append (nn.MaxPool2d (kernel_size=2) Stride=2) return nn.Sequential (* layers) VGG network

Like AlexNet and LeNet, VGG network can be divided into two parts: the first part is mainly composed of convolution layer and aggregation layer, and the second part is composed of full connection layer. As shown in the following figure:

The VGG neural network continuously connects several VGG blocks in the figure above (defined in the vgg_block function). There is a hyperparameter variable conv_arch. This variable specifies the number of convolution layers and the number of output channels in each VGG block. The full connection module is the same as in AlexNet.

There are five convolution blocks in the original VGG network, the first two blocks each have a convolution layer, and the last three blocks contain two convolution layers. The first module has 64 output channels, and each subsequent module doubles the number of output channels until the number reaches 512. Because the network uses eight convolution layers and three fully connected layers, it is often referred to as VGG-11.

Conv_arch = (1,64), (1,128), (2,256), (2,512), (2,512))

The following code implements VGG-11. This can be done simply by executing a for loop on the conv_arch.

Def vgg (conv_arch): conv_blks = [] in_channels = 1 # convolution layer part for (num_convs, out_channels) in conv_arch: conv_blks.append (vgg_block (num_convs, in_cannels, out_channels)) in_channels = out_channels return nn.Sequential (* conv_blks, nn.Flatten () # fully connected layer part nn.Linear (out_channels * 7 * 7, 0.5), nn.ReLU (), nn.Dropout (4096), nn.Linear (4096, 4096), nn.ReLU (), nn.Dropout (4096), nn.Linear (4096, 10) net = vgg (conv_arch)

Next, we will build a single-channel data sample with a height and width of 224 to observe the shape of the output of each layer.

X = torch.randn (size= (1,1,224,224) for blk in net: X = blk (X) print (blk.__class__.__name__, 'output shape:\ tasking, X.shape) Sequential output shape: torch.Size ([1,64,112,112]) Sequential output shape: torch.Size ([1,128,56,56]) Sequential output shape: torch.Size ([1,256,28,28]) Sequential output shape: torch.Size ([1] 512,1414]) Sequential output shape: torch.Size ([1,512,7,7]) Flatten output shape: torch.Size ([1, 25088]) Linear output shape: torch.Size ([1, 4096]) ReLU output shape: torch.Size ([1, 4096]) Dropout output shape: torch.Size ([1, 4096]) Linear output shape: torch.Size ([1, 4096]) ReLU output shape: torch.Size ([1, 4096]) Dropout output shape: torch.Size 4096]) Linear output shape: torch.Size ([1,10])

We halve the height and width of each block, resulting in a height and width of 7. Finally, it is flattened and fed into the fully connected layer for processing.

Training model

Because VGG-11 has more computation than AlexNet, we build a network with fewer channels, which is enough to train Fashion-MNIST data sets.

Ratio = 4small_conv_arch = [(pair [0], pair [1] / / ratio) for pair in conv_arch] net = vgg (small_conv_arch)

Except for using a slightly higher learning rate, the model training process is similar to that of AlexNet.

Lr, num_epochs, batch_size = 0.05,10, 128train_iter, test_iter = d2l.load_data_fashion_mnist (batch_size, resize=224) d2l.train_ch7 (net, train_iter, test_iter, num_epochs, lr, d2l.try_gpu () loss 0.175, train axx 0.935, test acc 0.9172559.6 examples/sec on cuda:0

This is the end of the content of "what is the pytorch network block VGG". Thank you for reading. If you want to know more about the industry, you can follow the website, the editor will output more high-quality practical articles for you!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.