7.7 DenseNet 稠密连接网络
1 原理
ResNet 将函数展开为 \(f(x) = x + g(x)\)。而 DenseNet 输出是连接而不是相加
左 ResNet 右 DenseNet
稠密网络主要由 2 部分构成：
- 稠密块：定义如何连接输入和输出
- 过渡层：控制通道数量，使其不会太复杂
2 稠密块
| import torch
from torch import nn
from d2l import torch as d2l
def conv_block(input_channels, num_channels):
return nn.Sequential(
nn.BatchNorm2d(input_channels),
nn.ReLU(),
nn.Conv2d(input_channels, num_channels, kernel_size=3, padding=1)
)
class DenseBlock(nn.Module):
def __init__(self, num_convs, input_channels, num_channels):
super(DenseBlock, self).__init__()
layer = []
for i in range(num_convs):
layer.append(conv_block(
num_channels * i + input_channels, num_channels))
self.net = nn.Sequential(*layer)
def forward(self, X):
for blk in self.net:
Y = blk(X)
# 连接通道维度上每个块的输入和输出
X = torch.cat((X, Y), dim=1)
return X
|
卷积块的通道数控制了输出通道数相对于输入通道数的增长，因此也被称为 增长率
| example |
|---|
| blk = DenseBlock(2, 3, 10)
X = torch.randn(4, 3, 8, 8)
Y = blk(X)
Y.shape # torch.Size([4, 23, 8, 8]) 23 = 3 + 2 * 10
|
3 过渡层
由于每个稠密块都会带来通道数的增加，使用过多则会过于复杂化模型。而过渡层可以用来控制模型复杂度。它通过 1 x 1 卷积层来减小通道数，并使用步幅为 2 的平均汇聚层减半高和宽，从而进一步降低模型复杂度
| def transition_block(input_channels, num_channels):
return nn.Sequential(
nn.BatchNorm2d(input_channels), nn.ReLU(),
nn.Conv2d(input_channels, num_channels, kernel_size=1),
nn.AvgPool2d(kernel_size=2, stride=2))
|
| example |
|---|
| blk = transition_block(23, 10)
blk(Y).shape # torch.Size([4, 10, 4, 4])
|
4 DenseNet 模型
| b1 = nn.Sequential(
nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
)
# 使用 4 个稠密块
# num_channels 为当前的通道数
num_channels, growth_rate = 64, 32
num_convs_in_dense_blocks = [4, 4, 4, 4]
blks = []
for i, num_convs in enumerate(num_convs_in_dense_blocks):
blks.append(DenseBlock(num_convs, num_channels, growth_rate))
# 上一个稠密块的输出通道数
num_channels += num_convs * growth_rate
# 在稠密块之间添加一个过渡层，使通道数量减半
if i != len(num_convs_in_dense_blocks) - 1:
blks.append(transition_block(num_channels, num_channels // 2))
num_channels = num_channels // 2
net = nn.Sequential(
b1, *blks,
nn.BatchNorm2d(num_channels),
nn.ReLU(),
nn.AdaptiveAvgPool2d((1, 1)),
nn.Flatten(),
nn.Linear(num_channels, 10)
)
|
将输入高和宽从 224 降到 96 来简化计算
| lr, num_epochs, batch_size = 0.1, 10, 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=96)
d2l.train_ch6(net, train_iter, test_iter, num_epochs, lr, d2l.try_gpu())
|