【Pytorch学习笔记】5.卷积神经网络

根据龙良曲Pytorch学习视频整理，视频链接：
【计算机-AI】PyTorch学这个就够了！
(好课推荐)深度学习与PyTorch入门实战——主讲人龙良曲

32.卷积神经网络

基础知识还是得看ng

import torch
import torch.nn as nn
import torch.nn.functional as F

# 卷积层
x = torch.rand(1, 1, 28, 28)

layer = nn.Conv2d(1, 3, kernel_size=3, stride=1, padding=0)
out = layer.forward(x)
print(out.size())   # torch.Size([1, 3, 26, 26])

layer = nn.Conv2d(1, 3, kernel_size=3, stride=1, padding=1)
out = layer.forward(x)
print(out.size())   # torch.Size([1, 3, 28, 28])

layer = nn.Conv2d(1, 3, kernel_size=3, stride=2, padding=1)
out = layer.forward(x)
print(out.size())   # torch.Size([1, 3, 14, 14])

out = layer(x)  # __call__ 推荐使用，不推荐使用forward
print(out.size())   # torch.Size([1, 3, 14, 14])

# print(layer.weight)
print(layer.weight.shape)   # torch.Size([3, 1, 3, 3])
print(layer.bias.shape)     # torch.Size([3])

w = torch.rand(16, 3, 5, 5)
b = torch.rand(16)
x = torch.randn(1, 3, 28, 28)
out = F.conv2d(x, w, b, stride=2, padding=2)
print(out.shape)    # torch.Size([1, 16, 14, 14])

33.池化层&上/下采样

Pooling：feature map变小，与隔行采样（Down sample）不同

• Max pooling
• Avg pooling

# 池化层
x = out
print(x.shape)      # torch.Size([1, 16, 14, 14])

layer = nn.MaxPool2d(2, stride=2)
out = layer(x)
print(out.shape)    # torch.Size([1, 16, 7, 7])

out = F.avg_pool2d(x, 2, stride=2)
print(out.shape)    # torch.Size([1, 16, 7, 7])

Up sample

# 上采样
x = out
out = F.interpolate(x, scale_factor=2, mode='nearest')
print(out.shape)    # torch.Size([1, 16, 14, 14])

ReLU

# ReLU
print(x.shape)      # torch.Size([1, 16, 7, 7])
layer = nn.ReLU(inplace=True)
out = layer(x)
print(out.shape)    # torch.Size([1, 16, 7, 7])

out = F.relu(x)
print(out.shape)    # torch.Size([1, 16, 7, 7])

34.批量正则化

1d

import torch
import torch.nn as nn

x = torch.randn(100, 16, 784)

layer = nn.BatchNorm1d(16, momentum=0.1, affine=True)   # affine自动更新beta、gama
# layer.eval()    # test时加上
out = layer(x)
print(layer.running_mean)
print(layer.running_var)

for i in range(100):
    out = layer(x)
print(layer.running_mean)
print(layer.running_var)
"""
tensor([-3.3207e-04, -2.9735e-04,  8.1316e-04,  9.3234e-05,  1.9049e-04,
         6.9931e-04,  2.9378e-04,  3.1153e-05,  3.4325e-04,  3.2283e-04,
        -2.0425e-04, -4.0346e-04,  1.7246e-04, -1.9482e-04, -1.2086e-04,
        -8.4132e-04])
tensor([0.9999, 0.9999, 1.0002, 1.0006, 0.9997, 1.0003, 1.0000, 1.0002, 1.0003,
        1.0000, 0.9998, 1.0000, 1.0005, 0.9996, 0.9997, 0.9999])
tensor([-0.0033, -0.0030,  0.0081,  0.0009,  0.0019,  0.0070,  0.0029,  0.0003,
         0.0034,  0.0032, -0.0020, -0.0040,  0.0017, -0.0019, -0.0012, -0.0084])
tensor([0.9986, 0.9987, 1.0023, 1.0065, 0.9969, 1.0033, 0.9997, 1.0022, 1.0029,
        1.0004, 0.9981, 0.9999, 1.0053, 0.9957, 0.9972, 0.9985])
"""

layer.running_mean和layer.running_var得到的是全局的均值和方差，不是当前Batch上的，第一次只跑了一个Batch，现在还没有办法直接查看某个Batch上的这两个统计量的值。第一次只进行一次前向传播，在前向传播中来更新均值和方差的值：$${\mu }^{\prime }=(1-m)\mu +m{\mu }_t=(1-0.1)\times 0+0.1\times 0.5=0.05$$
默认动量m = 0.1，$\mu$是更新前的均值（初始值为0），${\mu }_t$是当前batch的平均值。进行多次前向传播，均值和方差就会趋于数据真实分布

注：test时要加上layer.eval()，test不进行反向传播，即$\beta$和$\gamma$不更新

2d

# 接卷积神经网络实例
print(x.shape)              # torch.Size([1, 16, 7, 7])

layer = nn.BatchNorm2d(16)
out = layer(x)
print(out.shape)            # torch.Size([1, 16, 7, 7])
print(layer.weight.shape)   # torch.Size([16])
print(layer.bias.shape)     # torch.Size([16])
print(vars(layer))          # layer的参数

Advantages

• Converge faster
• Better performance
• Robust
  stable
  larger learning rate

35.经典卷积网络

ImageNet: LeNet-5 $\to$ AlexNet $\to$ VGG $\to$ GoogLeNet $\to$ ResNet $\to$ Inception

36.残差网络

import torch
import torch.nn as nn
import torch.nn.functional as F

class ResBlk(nn.Module):
    def __init__(self, ch_in, ch_out):
        self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn1 = nn.BatchNorm2d(ch_out)
        self.conv2 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(ch_out)
        
        self.extra = nn.Sequential()
        if ch_out != ch_in:
            # [b, ch_in, h, w] => [b, ch_out, h, w]
            self.extra = nn.Sequential(
                nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=1),
                nn.BatchNorm2d(ch_out)
            )
            
    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out = self.extra(x) + out
        return out

37.nn.Module

• Every Layer is nn.Module
• nn.Module nested in nn.Module

1. embed current layers
   Linear ReLU Sigmoid Conv2d ConvTransposed2d Dropout etc.
2. Container
   net(x)
3. parameters
4. modules
   modules: all nodes
   children: direct childrend
5. to(device)
6. save and load
   net.load_state_dict(torch.load('ckpt.mdl'))
torch.save(net.state_dict(), 'ckpt.mdl')
7. train / test
   net.train()
net.eval()
8. implement own layer

class Flatten(nn.Module):
    def __init__(self):
        super(Flatten, self).__init__()

    def forward(self, input):
        return input.view(input.size(0), -1)

class TestNet(nn.Module):
    def __init__(self):
        super(TestNet, self).__init__()
        self.net = nn.Sequential(nn.Conv2d(1, 16, stride=1, padding=1),
                                 nn.MaxPool2d(2, 2),
                                 Flatten(),
                                 nn.Linear(1*14*14, 10))
        
    def forward(self, x):
        return self.net(x)

38.数据增强

Limited Data

• Small network capacity
• Regularization
• Data argumentation（will help but not much）

Data argumentation

• Flip
  transforms.RandomHorizontalFlip() 随机水平翻转
  transforms.RandomVertialFlip() 随机垂直翻转
• Rotate
  transforms.RandomRotation(15) 随机旋转，不超过15度
  transforms.RandomRotation([90, 180, 270]) 随机在指定角度中旋转
• Scale
  transforms.Resize([32, 32])
• Random Move & Crop
  transforms.RandomCrop([28, 28])
• Noise
• GAN

39.实战

main.py

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import transforms, datasets
from Pytorch21_7_29.conv.lenet5 import Lenet5
from Pytorch21_7_29.conv.resnet import ResNet18

def main():
    batch_size = 32

    cifar_train = datasets.CIFAR10('cifar', train=True, transform=transforms.Compose([
        transforms.Resize([32, 32]),
        transforms.ToTensor()
    ]), download=True)
    cifar_train = DataLoader(cifar_train, batch_size=batch_size, shuffle=True)

    cifar_test = datasets.CIFAR10('cifar', train=False, transform=transforms.Compose([
        transforms.Resize([32, 32]),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ]), download=True)
    cifar_test = DataLoader(cifar_test, batch_size=batch_size, shuffle=True)

    x, label = iter(cifar_train).next()
    print("x:", x.shape, 'label:', label.shape) # x: torch.Size([32, 3, 32, 32]) label: torch.Size([32])

    device = torch.device('cuda')
    # model = Lenet5().to(device)
    model = ResNet18().to(device)
    criteon = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=1e-3)
    print(model)

    for epoch in range(1000):

        model.train()
        for batch_idx, (x, label) in enumerate(cifar_train):
            x, label = x.to(device), label.to(device)

            logits = model(x)
            # logits: [b, 10]   label: [b]
            loss = criteon(logits, label)   # tensor scalar

            # backward
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        #
        print(epoch, loss.item())

        # 测试不需要构建计算图
        model.eval()
        with torch.no_grad():
            # test
            total_correct = 0
            total_num = 0
            for x, label in cifar_test:
                x, label = x.to(device), label.to(device)
                logits = model(x)
                pred = logits.argmax(dim=1)
                total_correct += torch.eq(pred, label).float().sum().item()
                total_num += x.size(0)

            acc = total_correct / total_num
            print(epoch, acc)

if __name__ == '__main__':
    main()

"""
Lenet5(
  (conv_unit): Sequential(
    (0): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))
    (1): AvgPool2d(kernel_size=2, stride=2, padding=0)
    (2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
    (3): AvgPool2d(kernel_size=2, stride=2, padding=0)
  )
  (fc_unit): Sequential(
    (0): Linear(in_features=400, out_features=120, bias=True)
    (1): ReLU()
    (2): Linear(in_features=120, out_features=84, bias=True)
    (3): ReLU()
    (4): Linear(in_features=84, out_features=10, bias=True)
  )
  (criteon): CrossEntropyLoss()
)

ResNet18(
  (conv1): Sequential(
    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(3, 3))
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (blk1): ResBlk(
    (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (extra): Sequential(
      (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2))
      (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (blk2): ResBlk(
    (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (extra): Sequential(
      (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2))
      (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (blk3): ResBlk(
    (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (extra): Sequential(
      (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2))
      (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (blk4): ResBlk(
    (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (extra): Sequential()
  )
  (outlayer): Linear(in_features=512, out_features=10, bias=True)
)
"""

lenet5.py

import torch
import torch.nn as nn
import torch.nn.functional as F

class Lenet5(nn.Module):
    """
    for cifar10 dataset.
    """
    def __init__(self):
        super(Lenet5, self).__init__()

        self.conv_unit = nn.Sequential(
            # x:[b, 3, 32, 32] => [b, 6, 28, 28]
            nn.Conv2d(3, 6, kernel_size=5, stride=1, padding=0),
            nn.AvgPool2d(kernel_size=2, stride=2, padding=0),
            # x:[b, 6, 14, 14] => [b, 16, 10, 10]
            nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0),
            nn.AvgPool2d(kernel_size=2, stride=2, padding=0),
            # x:[b, 16, 5, 5] => [b, 400]
        )
        # flatten
        # fc unit
        self.fc_unit = nn.Sequential(
            nn.Linear(16*5*5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 10)
        )

        # self.criteon = nn.MSELoss()           # logistics问题使用
        self.criteon = nn.CrossEntropyLoss()    # 一般分类问题使用

    def forward(self, x):
        """
        :param x: [b, 3, 32, 32]
        :return:
        """
        batch_size = x.size(0)
        x = self.conv_unit(x)
        x = x.view(batch_size, -1)  # -1指16*5*5
        logits = self.fc_unit(x)

        # pred = F.softmax(logits, dim=1)   交叉熵函数内包含了softmax
        # loss = self.criteon(logits, y)

        return logits


def main():
    net = Lenet5()
    tmp = torch.randn(2, 3, 32, 32)
    out = net(tmp)
    print('lenet_out:', out.shape)  # lenet_out: torch.Size([2, 10])

if __name__ == '__main__':
    main()

resnet.py

import torch
import torch.nn as nn
import torch.nn.functional as F

class ResBlk(nn.Module):
    """
    resnet block
    """

    def __init__(self, ch_in, ch_out, stride=1):
        """
        :param ch_in:
        :param ch_out:
        """
        super(ResBlk, self).__init__()

        # we add stride support for resblk, which is distinct from tutorials.
        self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(ch_out)
        self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(ch_out)

        self.extra = nn.Sequential()
        if ch_out != ch_in:
            # [b, ch_in, h, w] => [b, ch_out, h, w]
            self.extra = nn.Sequential(
                nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride),
                nn.BatchNorm2d(ch_out)
            )

    def forward(self, x):
        """
        :param x: [b, ch, h, w]
        :return:
        """
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        # short cut
        out = self.extra(x) + out
        out = F.relu(out)
        return out

class ResNet18(nn.Module):

    def __init__(self):
        super(ResNet18, self).__init__()

        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),
            nn.BatchNorm2d(64)
        )

        # followed 4 blocks
        # [b, 64, h, w] => [b, 128, h, w]
        self.blk1 = ResBlk(64, 128, stride=2)
        # [b, 128, h, w] => [b, 256, h, w]
        self.blk2 = ResBlk(128, 256, stride=2)
        # [b, 256, h, w] => [b, 512, h, w]
        self.blk3 = ResBlk(256, 512, stride=2)
        # [b, 512, h, w] => [b, 1024, h, w]
        self.blk4 = ResBlk(512, 512, stride=2)

        self.outlayer = nn.Linear(512*1*1, 10)

    def forward(self, x):
        """
        :param x:
        :return:
        """
        x = F.relu(self.conv1(x))

        x = self.blk1(x)
        x = self.blk2(x)
        x = self.blk3(x)
        x = self.blk4(x)
        # print("after_conv:", x.shape) # torch.Size([32, 512, 2, 2])
        # [b, 512, h, w] => [b, 512, 1, 1]
        x = F.adaptive_avg_pool2d(x, [1, 1])
        # print("after_pool:", x.shape) # torch.Size([32, 512, 1, 1])
        x = x.view(x.size(0), -1)

        x = self.outlayer(x)

        return x

def main():
    blk = ResBlk(64, 128, stride=2)
    tmp = torch.randn(2, 64, 32, 32)
    out = blk(tmp)
    print("block:", out.shape)      # torch.Size([2, 128, 16, 16])

    x = torch.randn(2, 3, 32, 32)
    model = ResNet18()
    out = model(x)
    print("resnet:", out.shape)     # torch.Size([2, 10])

if __name__ == '__main__':
    main()