（pytorch-深度学习）实现稠密连接网络（DenseNet）

稠密连接网络（DenseNet）

ResNet中的跨层连接设计引申出了数个后续工作。稠密连接网络（DenseNet）与ResNet的主要区别在于在跨层连接上的主要区别：

• ResNet使用相加
• DenseNet使用连结

ResNet（左）与DenseNet（右）:

图中将部分前后相邻的运算抽象为模块$A$和模块$B$。

• DenseNet里模块$B$的输出不是像ResNet那样和模块$A$的输出相加，而是在通道维上连结。
• 这样模块$A$的输出可以直接传入模块$B$后面的层。在这个设计里，模块$A$相当于直接跟模块$B$后面的所有层直接连接在了一起。这也是它被称为“稠密连接”的原因。

DenseNet的主要构建模块是稠密块（dense block）和过渡层（transition layer）。

• 稠密块定义了输入和输出是如何连结的
• 过渡层用来控制通道数，控制其大小

稠密块

DenseNet使用了ResNet改良版的“批量归一化、激活和卷积”结构：

import time
import torch
from torch import nn, optim
import torch.nn.functional as F

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def conv_block(in_channels, out_channels):
    blk = nn.Sequential(nn.BatchNorm2d(in_channels), 
                        nn.ReLU(),
                        nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1))
    return blk

• 稠密块由多个conv_block组成，每块使用相同的输出通道数。
• 在前向计算时，我们将每块的输入和输出在通道维上连结。

class DenseBlock(nn.Module):
    def __init__(self, num_convs, in_channels, out_channels):
        super(DenseBlock, self).__init__()
        net = []
        for i in range(num_convs):
            in_c = in_channels + i * out_channels
            net.append(conv_block(in_c, out_channels))
        self.net = nn.ModuleList(net)
        self.out_channels = in_channels + num_convs * out_channels # 计算输出通道数

    def forward(self, X):
        for blk in self.net:
            Y = blk(X)
            X = torch.cat((X, Y), dim=1)  # 在通道维上将输入和输出连结
        return X

定义一个有2个输出通道数为10的卷积块。

• 使用通道数为3的输入时，我们会得到通道数为$3+2\times 10=23$的输出。
• 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（growth rate）。

blk = DenseBlock(2, 3, 10)
X = torch.rand(4, 3, 8, 8)
Y = blk(X)
Y.shape # torch.Size([4, 23, 8, 8])

过渡层

• 每个稠密块都会带来通道数的增加，使用过多则会带来过于复杂的模型。
• 过渡层用来控制模型复杂度。它通过$1\times 1$卷积层来减小通道数，并使用步幅为2的平均池化层减半高和宽，从而进一步降低模型复杂度。

def transition_block(in_channels, out_channels):
    blk = nn.Sequential(
            nn.BatchNorm2d(in_channels), 
            nn.ReLU(),
            nn.Conv2d(in_channels, out_channels, kernel_size=1),
            nn.AvgPool2d(kernel_size=2, stride=2))
    return blk

对上例中稠密块的输出，使用通道数为10的过渡层。此时输出的通道数减为10，高和宽均减半。

blk = transition_block(23, 10)
blk(Y).shape # torch.Size([4, 10, 4, 4])

DenseNet模型

DenseNet首先使用和ResNet一样的单卷积层和最大池化层。

net = 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))

• 接着使用4个稠密块。
• 同ResNet一样，我们可以设置每个稠密块使用多少个卷积层（这里设成4）。
• 稠密块里的卷积层通道数（即增长率）设为32，所以每个稠密块将增加128个通道。

ResNet里通过步幅为2的残差块在每个模块之间减小高和宽。DenseNet则使用过渡层来减半高和宽，并减半通道数。

num_channels, growth_rate = 64, 32  # num_channels为当前的通道数
num_convs_in_dense_blocks = [4, 4, 4, 4]

for i, num_convs in enumerate(num_convs_in_dense_blocks):
    DB = DenseBlock(num_convs, num_channels, growth_rate)
    net.add_module("DenseBlosk_%d" % i, DB)
    # 上一个稠密块的输出通道数
    num_channels = DB.out_channels
    # 在稠密块之间加入通道数减半的过渡层
    if i != len(num_convs_in_dense_blocks) - 1:
        net.add_module("transition_block_%d" % i, transition_block(num_channels, num_channels // 2))
        num_channels = num_channels // 2

• 最后接上全局池化层和全连接层来输出。

class GlobalAvgPool2d(nn.Module):
    # 全局平均池化层可通过将池化窗口形状设置成输入的高和宽实现
    def __init__(self):
        super(GlobalAvgPool2d, self).__init__()
    def forward(self, x):
        return F.avg_pool2d(x, kernel_size=x.size()[2:])

class FlattenLayer(torch.nn.Module):
    def __init__(self):
        super(FlattenLayer, self).__init__()
    def forward(self, x): # x shape: (batch, *, *, ...)
        return x.view(x.shape[0], -1)

net.add_module("BN", nn.BatchNorm2d(num_channels))
net.add_module("relu", nn.ReLU())
net.add_module("global_avg_pool", GlobalAvgPool2d()) # GlobalAvgPool2d的输出: (Batch, num_channels, 1, 1)
net.add_module("fc", nn.Sequential(FlattenLayer(), nn.Linear(num_channels, 10))) 

• 打印每个子模块的输出维度

X = torch.rand((1, 1, 96, 96))
for name, layer in net.named_children():
    X = layer(X)
    print(name, ' output shape:\t', X.shape)

0  output shape:	 torch.Size([1, 64, 48, 48])
1  output shape:	 torch.Size([1, 64, 48, 48])
2  output shape:	 torch.Size([1, 64, 48, 48])
3  output shape:	 torch.Size([1, 64, 24, 24])
DenseBlosk_0  output shape:	 torch.Size([1, 192, 24, 24])
transition_block_0  output shape:	 torch.Size([1, 96, 12, 12])
DenseBlosk_1  output shape:	 torch.Size([1, 224, 12, 12])
transition_block_1  output shape:	 torch.Size([1, 112, 6, 6])
DenseBlosk_2  output shape:	 torch.Size([1, 240, 6, 6])
transition_block_2  output shape:	 torch.Size([1, 120, 3, 3])
DenseBlosk_3  output shape:	 torch.Size([1, 248, 3, 3])
BN  output shape:	 torch.Size([1, 248, 3, 3])
relu  output shape:	 torch.Size([1, 248, 3, 3])
global_avg_pool  output shape:	 torch.Size([1, 248, 1, 1])
fc  output shape:	 torch.Size([1, 10])

• 获取数据

def load_data_fashion_mnist(batch_size, resize=None, root='~/Datasets/FashionMNIST'):
    """Download the fashion mnist dataset and then load into memory."""
    trans = []
    if resize:
        trans.append(torchvision.transforms.Resize(size=resize))
    trans.append(torchvision.transforms.ToTensor())
    
    transform = torchvision.transforms.Compose(trans)
    mnist_train = torchvision.datasets.FashionMNIST(root=root, train=True, download=True, transform=transform)
    mnist_test = torchvision.datasets.FashionMNIST(root=root, train=False, download=True, transform=transform)
    if sys.platform.startswith('win'):
        num_workers = 0  # 0表示不用额外的进程来加速读取数据
    else:
        num_workers = 4
    train_iter = torch.utils.data.DataLoader(mnist_train, batch_size=batch_size, shuffle=True, num_workers=num_workers)
    test_iter = torch.utils.data.DataLoader(mnist_test, batch_size=batch_size, shuffle=False, num_workers=num_workers)

    return train_iter, test_iter

batch_size = 256
# 如出现“out of memory”的报错信息，可减小batch_size或resize
train_iter, test_iter = load_data_fashion_mnist(batch_size, resize=96)

训练模型

def train(net, train_iter, test_iter, batch_size, optimizer, device, num_epochs):
    net = net.to(device)
    print("training on ", device)
    loss = torch.nn.CrossEntropyLoss()
    for epoch in range(num_epochs):
        train_l_sum, train_acc_sum, n, batch_count, start = 0.0, 0.0, 0, 0, time.time()
        for X, y in train_iter:
            X = X.to(device)
            y = y.to(device)
            y_hat = net(X)
            l = loss(y_hat, y)
            optimizer.zero_grad()
            l.backward()
            optimizer.step()
            train_l_sum += l.cpu().item()
            train_acc_sum += (y_hat.argmax(dim=1) == y).sum().cpu().item()
            n += y.shape[0]
            batch_count += 1
        test_acc = evaluate_accuracy(test_iter, net)
        print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f, time %.1f sec'
              % (epoch + 1, train_l_sum / batch_count, train_acc_sum / n, test_acc, time.time() - start))

lr, num_epochs = 0.001, 5
optimizer = torch.optim.Adam(net.parameters(), lr=lr)
train(net, train_iter, test_iter, batch_size, optimizer, device, num_epochs)

《动手学深度学习》