视频指路
参考博客笔记
参考笔记二

上课笔记

可以设置padding=‘same’ 使输入输出大小一致

10.1GoogleNet（Inception 层）

说明：Inception Moudel

1、卷积核超参数选择困难（提供四条变换路线输出要保证宽高一致，把结果concatenate到一起，效率高的权重大），自动找到卷积的最佳组合。

2、1x1卷积核，不同通道的信息融合。使用1x1卷积核虽然参数量增加了，但是能够显著的降低计算量(operations)

3、Inception Moudel由4个分支组成，要分清哪些是在Init里定义，哪些是在forward里调用。4个分支在dim=1(channels)上进行concatenate。24+16+24+24 = 88

4、GoogleNet的Inception(Pytorch实现)

下面是1*1卷积核计算过程

在5 *5 的卷积之前先进行一个1 * 1的卷积能有效降低运算量

比如：192 * 28 * 28经过一个5 * 5的卷积得到 32 * 28 * 28的输出运算为：

5^2 * 28 ^2 * 192 * 32=120422400

而中间先经过一个1 * 1的卷积再经过一个5 * 5的卷积得到 32 * 28 * 28的输出运算为：1^2 * 28^2 * 192 * 16 + 5^2 + 28^2 * 16 * 32 = 12433648

少了十倍

1*1卷积的主要作用有以下几点：

1、降维。比如，一张500 * 500且厚度depth为100 的图片在20个filter上做11的卷积，那么结果的大小为500500*20。

2、加入非线性。卷积层之后经过激励层，1*1的卷积在前一层的学习表示上添加了非线性激励，提升网络的表达能力；

3、增加模型深度。可以减少网络模型参数，增加网络层深度，一定程度上提升模型的表征能力。

self.branch1x1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch5x5_1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch5x5_2 = nn.Conv2d(16, 24, kernel_size=5, padding=2)self.branch3x3_1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch3x3_2 = nn.Conv2d(16, 24, kernel_size=3, padding=1)self.branch3x3_3 = nn.Conv2d(24, 24, kernel_size=3, padding=1)self.branch_pool = nn.Conv2d(in_channels, 24, kernel_size=1)

把上面四个卷积出的通道聚合(Concatenate)，输出

outputs = [branch1*1, branch5*5, branch3*3, branch_pool]`
return torch.cat(outputs, dim=1)#沿着通道c拼接起来 维度=1

张量的维度是(b, c, h, w) batch, channel, width, height

代码实现

import torch
import torch.nn as nn
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.nn.functional as F
import torch.optim as optim# prepare datasetbatch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])  # 归一化,均值和方差train_dataset = datasets.MNIST(root='../dataset/mnist/', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
test_dataset = datasets.MNIST(root='../dataset/mnist/', train=False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size)# design model using class
class InceptionA(nn.Module):def __init__(self, in_channels):super(InceptionA, self).__init__()self.branch1x1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch5x5_1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch5x5_2 = nn.Conv2d(16, 24, kernel_size=5, padding=2)self.branch3x3_1 = nn.Conv2d(in_channels, 16, kernel_size=1)self.branch3x3_2 = nn.Conv2d(16, 24, kernel_size=3, padding=1)self.branch3x3_3 = nn.Conv2d(24, 24, kernel_size=3, padding=1)self.branch_pool = nn.Conv2d(in_channels, 24, kernel_size=1)def forward(self, x):branch1x1 = self.branch1x1(x)branch5x5 = self.branch5x5_1(x)branch5x5 = self.branch5x5_2(branch5x5)branch3x3 = self.branch3x3_1(x)branch3x3 = self.branch3x3_2(branch3x3)branch3x3 = self.branch3x3_3(branch3x3)branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)branch_pool = self.branch_pool(branch_pool)outputs = [branch1x1, branch5x5, branch3x3, branch_pool]return torch.cat(outputs, dim=1)  # b,c,w,h  c对应的是dim=1class Net(nn.Module):def __init__(self):super(Net, self).__init__()self.conv1 = nn.Conv2d(1, 10, kernel_size=5)self.conv2 = nn.Conv2d(88, 20, kernel_size=5)  # 88 = 24x3 + 16self.incep1 = InceptionA(in_channels=10)  # 与conv1 中的10对应self.incep2 = InceptionA(in_channels=20)  # 与conv2 中的20对应self.mp = nn.MaxPool2d(2)self.fc = nn.Linear(1408, 10)def forward(self, x):in_size = x.size(0)x = F.relu(self.mp(self.conv1(x)))x = self.incep1(x)x = F.relu(self.mp(self.conv2(x)))x = self.incep2(x)x = x.view(in_size, -1)x = self.fc(x)return xmodel = Net()# construct loss and optimizer
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)# training cycle forward, backward, updatedef train(epoch):running_loss = 0.0for batch_idx, data in enumerate(train_loader, 0):inputs, target = dataoptimizer.zero_grad()outputs = model(inputs)loss = criterion(outputs, target)loss.backward()optimizer.step()running_loss += loss.item()if batch_idx % 300 == 299:print('[%d, %5d] loss: %.3f' % (epoch + 1, batch_idx + 1, running_loss / 300))running_loss = 0.0def test():correct = 0total = 0with torch.no_grad():for data in test_loader:images, labels = dataoutputs = model(images)_, predicted = torch.max(outputs.data, dim=1)total += labels.size(0)correct += (predicted == labels).sum().item()print('accuracy on test set: %。3f %% ' % (100 * correct / total))if __name__ == '__main__':for epoch in range(10):train(epoch)test()

10.2 Residual Net

知识点：残差层定义

问题描述：卷积核层数不是越深越好，可能存在梯度消失

主要思路：引入残差连接，拼接后再激活，计算梯度的时候就能有所保留，要求输入输出大小相同

代码实现：定义残差块类，指定输入通道数，跳转拼接后再激活。模型构建时再定义相关层

跳连接：将H(x)的输入再加一个x，求导的时候x`=1，那么就算梯度很小也是将近于1，多个这样的数相乘梯度还是不为0，能解决梯度消失的情况，其中F(x)和x应该尺寸相同。

class ResidualBlock(torch.nn.Module):def __init__(self, channels):super(ResidualBlock,self).__init__()self.channels = channels  # 过残差连接输入输出通道不变self.conv1 = torch.nn.Conv2d(channels, channels, 3,padding=1)#padding=1使得F(x)和x应该尺寸相同self.conv2 = torch.nn.Conv2d(channels, channels, 3,padding=1)def forward(self, x):y = F.relu(self.conv1(x))y = self.conv2(y)return F.relu(x + y)  # 先求和再激活

1.看一些深度学习理论方面的书比如花书

2.阅读pytorch文档，至少通读一遍

3.复现经典工作：先读代码，训练架构，测试架构，数据读取架构，损失函数怎么构建的，根据论文讲的东西自己去写

4.选一个特定领域阅读大量论文，看一下大家在设计网络的时候都用了什么技巧，想创新点

代码实现

import torch
from torchvision import datasets
from torch.utils.data import DataLoader
from torchvision import transforms
import torch.nn.functional as F# 1.数据集准备
batch_size = 64
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
train_dataset = datasets.MNIST(root='../dataset/minist', train = True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=2)
test_dataset = datasets.MNIST(root='../dataset/minist', train = False, download=True, transform=transform)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=batch_size, num_workers=2)# 2.模型构建
class ResidualBlock(torch.nn.Module):def __init__(self, channels):super(ResidualBlock,self).__init__()self.channels = channels  # 过残差连接输入输出通道不变self.conv1 = torch.nn.Conv2d(channels, channels, 3,padding=1)self.conv2 = torch.nn.Conv2d(channels, channels, 3,padding=1)def forward(self, x):y = F.relu(self.conv1(x))y = self.conv2(y)return F.relu(x + y)  # 先求和再激活class Net(torch.nn.Module):def __init__(self):super(Net, self).__init__()self.c1 = torch.nn.Conv2d(1, 16, 5)self.c2 = torch.nn.Conv2d(16, 32, 5)self.rblock1 = ResidualBlock(16)  # 与conv1 中的16对应self.rblock2 = ResidualBlock(32)  # 与conv2 中的32对应self.mp = torch.nn.MaxPool2d(2)  # 图像缩小一半 12  （不要改步长啊）self.l = torch.nn.Linear(512, 10)def forward(self, x):batch = x.size(0)x = torch.relu(self.mp(self.c1(x)))  # b*10*12*12x = self.rblock1(x)x = torch.relu(self.mp(self.c2(x)))  # b*20*4*4x = self.rblock2(x)x = x.view(batch, -1)# print(x.shape)return self.l(x)device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net().to(device)
# # 创建一个示例图像查看模型输出shape（好给全连接层赋值）----->输出：torch.Size([1, 1408])
# sample_image = torch.randn(1, 1, 28, 28)  # 1张图像，1个通道，28x28大小的图像
# output = model(sample_image)# 3.损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)# 4.训练
def train(epoch):running_loss = 0.0for batch_idex, (x, y) in enumerate(train_loader):x, y = x.to(device), y.to(device)y_pred = model(x)optimizer.zero_grad()loss = criterion(y_pred, y)running_loss += loss.item()loss.backward()optimizer.step()if batch_idex % 300 == 299:print(f'epoch{epoch+1}--------batch{batch_idex+1}-------loss={round(running_loss/300, 3)}')running_loss = 0.0def test():total = 0acc = 0with torch.no_grad():for (x, y) in test_loader:x, y = x.to(device), y.to(device)y_pred = model(x)total += y_pred.size(0)_, predicted = torch.max(y_pred, dim=1)acc += (predicted == y).sum().item()print('test= %.3f %%' % (100 * acc/total))if __name__ == '__main__':for epoch in range(10):train(epoch)test()

练习：阅读Identity Mappings in Deep Residual Networks，Densely Connected Convolutional Networks，实现相关网络用minist数据集测试