LetNet、Alex、VggNet分析及其pytorch實現

簡單分析一下主流的幾種神經網路

LeNet

LetNet作為卷積神經網路中的HelloWorld,它的結構及其的簡單，1998年由LeCun提出

基本過程:

可以看到LeNet-5跟現有的conv->pool->ReLU的套路不同，它使用的方式是conv1->pool->conv2->pool2再接全連線層，但是不變的是，卷積層後緊接池化層的模式依舊不變。

程式碼：

import torch.nn as nn
import torch


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

        layer1 = nn.Sequential()
        # Convolution with 5x5 kernel+2padding:28×28×6
        layer1.add_module('conv1', nn.Conv2d(1, 6, kernel_size=(3, 3), padding=1))
        # Pool with 2x2 average kernel+2 stride:14×14×6
        layer1.add_module('pool1', nn.MaxPool2d(kernel_size=2))
        self.layer1 = layer1

        layer2 = nn.Sequential()
        # Convolution with 5x5 kernel (no pad):10×10×16
        layer2.add_module('conv2', nn.Conv2d(6, 16, kernel_size=(5, 5)))
        # Pool with 2x2 average kernel+2 stride: 5x5×16
        layer2.add_module('pool2', nn.MaxPool2d(kernel_size=2))
        self.layer2 = layer2

        layer3 = nn.Sequential()
        # 5 = ((28/2)-4)/2
        layer3.add_module('fc1', nn.Linear(16 * 5 * 5, 120))
        layer3.add_module('fc2', nn.Linear(120, 84))
        layer3.add_module('fc3', nn.Linear(84, 10))
        self.layer3 = layer3

    def forward(self, x):
        x = self.layer1(x)
        # print(x.size())
        x = self.layer2(x)
        # print(x.size())
        # 展平x
        x = torch.flatten(x, 1)
        x = self.layer3(x)
        return x


# 測試
test_data = torch.rand(1, 1, 28, 28)
model = LeNet()
model(test_data)

輸出

tensor([[ 0.0067, -0.0431,  0.1072,  0.1275,  0.0143,  0.0865, -0.0490, -0.0936,
         -0.0315, -0.0367]], grad_fn=<AddmmBackward0>)

AlexNet

這個圖看起來稍微可能有億點複雜，其實這個是因為當時的GPU計算能力不太行，而Alex又比較複雜，所以Alex使用了兩個GPU並行來做運算，現在已經完全可以用一個GPU來代替了。
相對於LeNet來說，Alex網路層數更深，同時第一次引入了啟用層ReLU，又在全連線層引入了Dropout層防止過擬合。
執行流程圖在上面，跟LeNet的執行流程圖放在一張圖上。

程式碼：

import torch.nn as nn
import torch


class AlexNet(nn.Module):
    def __init__(self, num_class):
        super(AlexNet, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4)),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(64, 192, kernel_size=(5, 5), padding=2),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(192, 384, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(384, 256, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(256, 256, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=3, stride=2),
        )

        self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(256*5*5, 4096),
            nn.ReLU(True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(True),
            nn.Linear(4096, num_class)
        )

    def forward(self, x):
        x = self.features(x)
        print(x.size())
        x = x.view(x.size(0), 256 * 5 * 5)
        x = self.classifier(x)
        return x

# 測試
test_data = torch.rand(1, 3, 224, 224)
model = AlexNet(10)
model(test_data)

輸出：

torch.Size([1, 256, 5, 5])
tensor([[-0.0044,  0.0114,  0.0032, -0.0099,  0.0035, -0.0024,  0.0103, -0.0194,
          0.0149,  0.0094]], grad_fn=<AddmmBackward0>)

VggNet

VggNet是ImageNet 2014年的亞軍，總的來說就是它使用了更小的濾波器，用了更深的結構來提升深度學習的效果，從圖裡面可以看出來這一點，它沒有使用11*11這麼大的濾波器，取而代之的使用的都是3*3這種小的濾波器，它之所以使用很多小的濾波器，是因為層疊很多小的濾波器的感受野和一個大的濾波器的感受野是相同的，還能減少引數。

程式碼實現：

import torch.nn as nn


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

        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(64, 64, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(64, 128, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(128, 128, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(128, 256, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(256, 256, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(256, 256, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(256, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
            nn.ReLU(True),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(512 * 7 * 7, 4096), nn.ReLU(True),
            nn.Dropout(),
            nn.Linear(4096, 4096), nn.ReLU(True),
            nn.Dropout(),
            nn.Linear(4096, num_class),
        )
        self._initialize_weights()

    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        x = self.classifier(x)
        return x

使用卷積神經網路實現對Minist資料集的預測

程式碼：

import matplotlib.pyplot as plt
import torch.utils.data
import torchvision.datasets
import os
import torch.nn as nn
from torchvision import transforms


class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.layer1 = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=(3, 3)),
            nn.BatchNorm2d(16),
            nn.ReLU(inplace=True),
        )

        self.layer2 = nn.Sequential(
            nn.Conv2d(16, 32, kernel_size=(3, 3)),
            nn.BatchNorm2d(32),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )

        self.layer3 = nn.Sequential(
            nn.Conv2d(32, 64, kernel_size=(3, 3)),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True)
        )

        self.layer4 = nn.Sequential(
            nn.Conv2d(64, 128, kernel_size=(3, 3)),
            nn.BatchNorm2d(128),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2)
        )

        self.fc = nn.Sequential(
            nn.Linear(128 * 4 * 4, 1024),
            nn.ReLU(inplace=True),
            nn.Linear(1024, 128),
            nn.Linear(128, 10)
        )

    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        return x



os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"

data_tf = transforms.Compose(
    [transforms.ToTensor(),
     transforms.Normalize([0.5], [0.5])]
)

train_dataset = torchvision.datasets.MNIST(root='F:/機器學習/pytorch/書/data/mnist', train=True,
                                           transform=data_tf, download=True)

test_dataset = torchvision.datasets.MNIST(root='F:/機器學習/pytorch/書/data/mnist', train=False,
                                          transform=data_tf, download=True)

batch_size = 100
train_loader = torch.utils.data.DataLoader(
    dataset=train_dataset, batch_size=batch_size
)

test_loader = torch.utils.data.DataLoader(
    dataset=test_dataset, batch_size=batch_size
)

model = CNN()
model = model.cuda()
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
optimizer = torch.optim.Adam(model.parameters())

# 節約時間，三次夠了
iter_step = 3
loss1 = []
loss2 = []
for step in range(iter_step):
    loss1_count = 0
    loss2_count = 0
    for images, labels in train_loader:
        images = images.cuda()
        labels = labels.cuda()
        images = images.reshape(-1, 1, 28, 28)
        output = model(images)
        pred = output.squeeze()

        optimizer.zero_grad()
        loss = criterion(pred, labels)
        loss.backward()
        optimizer.step()

        _, pred = torch.max(pred, 1)

        loss1_count += int(torch.sum(pred == labels)) / 100
# 測試
    else:
        test_loss = 0
        accuracy = 0
        with torch.no_grad():
            for images, labels in test_loader:
                images = images.cuda()
                labels = labels.cuda()
                pred = model(images.reshape(-1, 1, 28, 28))
                _, pred = torch.max(pred, 1)
                loss2_count += int(torch.sum(pred == labels)) / 100

    loss1.append(loss1_count / len(train_loader))
    loss2.append(loss2_count / len(test_loader))

    print(f'第{step}次訓練：訓練準確率：{loss1[len(loss1)-1]}，測試準確率：{loss2[len(loss2)-1]}')

plt.plot(loss1, label='Training loss')
plt.plot(loss2, label='Validation loss')
plt.legend()

輸出：

第0次訓練：訓練準確率：0.9646166666666718，測試準確率：0.9868999999999996
第1次訓練：訓練準確率：0.9865833333333389，測試準確率：0.9908999999999998
第2次訓練：訓練準確率：0.9917000000000039，測試準確率：0.9879999999999994
<matplotlib.legend.Legend at 0x21f03092fd0>