ResNet¶
Note
ResNet的残差块加了一个恒等映射通路，使得我们能够训练更深的网络
ResNet是2015年ImageNet比赛的冠军
结构¶
下图比较了一个正常块和一个残差块，残差块加了一个恒等映射通路，在最后的激活函数前相加：
残差块使得随着网络的加深，函数空间总是嵌套的！而且它能缓解反向传播时的梯度消失和梯度爆炸，最终让ResNet又快又好：
在GoogleNet节中我们知道，1×1 卷积可以调整通道数和分辨率，它的功能类似于恒等映射，因此我们可以做出两类残差块：
import torch
from torch import nn
import torch.nn.functional as F
class Residual(nn.Module):
"""ResNet的残差块"""
def __init__(self, input_channels, num_channels, use_1x1conv=False,
strides=1):
# use_1x1conv=False时，必须input_channels=num_channels & strides=1
# 不然相加时shap不一致
super().__init__()
# 第一个卷积层完成通道和分辨率转换
self.conv1 = nn.Conv2d(input_channels, num_channels, kernel_size=3,
padding=1, stride=strides)
self.conv2 = nn.Conv2d(num_channels, num_channels, kernel_size=3,
padding=1)
# 使用1*1卷积完成通道和分辨率转换
if use_1x1conv:
self.conv3 = nn.Conv2d(input_channels, num_channels,
kernel_size=1, stride=strides)
else:
self.conv3 = None
# BatchNorm在卷积层和激活函数之间
self.bn1 = nn.BatchNorm2d(num_channels)
self.bn2 = nn.BatchNorm2d(num_channels)
def forward(self, X):
# 正常块
Y = F.relu(self.bn1(self.conv1(X)))
Y = self.bn2(self.conv2(Y))
# 1*1 conv layer instead of identity sometimes
if self.conv3:
X = self.conv3(X)
Y += X
return F.relu(Y)
实现¶
下面实现我们简易版的ResNet。
# 开头为普通卷积层
b1 = nn.Sequential(nn.Conv2d(1, 64, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(64), nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
# 第一个残差模块
b2 = nn.Sequential(Residual(64, 64),
Residual(64, 128, use_1x1conv=True, strides=2),
Residual(128, 128))
# 第二个残差模块
b3 = nn.Sequential(Residual(128, 256, use_1x1conv=True, strides=2),
Residual(256, 256))
# ResNet18
net = nn.Sequential(b1, b2, b3, nn.AdaptiveAvgPool2d((1, 1)),
nn.Flatten(), nn.Linear(256, 10))
X = torch.rand(size=(1, 1, 28, 28))
# 打印各部分的shape
for layer in net:
X = layer(X)
print(layer.__class__.__name__, 'output shape:\t', X.shape)
Sequential output shape: torch.Size([1, 64, 14, 14])
Sequential output shape: torch.Size([1, 128, 7, 7])
Sequential output shape: torch.Size([1, 256, 4, 4])
AdaptiveAvgPool2d output shape: torch.Size([1, 256, 1, 1])
Flatten output shape: torch.Size([1, 256])
Linear output shape: torch.Size([1, 10])
训练¶
import d2l
# 载入数据
batch_size = 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size=batch_size)
# 训练
lr, num_epochs = 0.01, 10
d2l.train_image_classifier(net, train_iter, test_iter, lr, num_epochs)
loss 0.141, train acc 0.947367, test acc 0.911000
