全連接限制圖像的尺寸弟孟，而卷積則不關(guān)心圖像尺寸大小汁咏，只需要接受輸入的通道數(shù)亚斋，輸出的通道數(shù)和卷積核大小即可確定圖像尺寸的變換過(guò)程，即
padding:對(duì)輸入圖片進(jìn)行填充攘滩，一般用0填充帅刊，padding=1，代表填充一圈漂问，保證卷積前后的圖像尺寸大小一致赖瞒，padding計(jì)算公式如下：
stride步長(zhǎng)：指的是卷積核每次滑動(dòng)的距離大小
本文采用GoogleNet來(lái)構(gòu)建深度網(wǎng)絡(luò)模型
1. 數(shù)據(jù)集構(gòu)建
每個(gè)像素點(diǎn)即每條數(shù)據(jù)中的值范圍為0-255，有的數(shù)字過(guò)大不利于訓(xùn)練且難以收斂蚤假，故將其歸一化到（0-1）之間
# 數(shù)據(jù)集處理
transform = transforms.Compose([
transforms.ToTensor(), # 轉(zhuǎn)化成Tensor張量
transforms.Normalize((0.1307,), (0.3081,)) # 歸一化處理,將其(0-255)映射到(0-1)
])
# 1.準(zhǔn)備數(shù)據(jù)集
train_dataset = datasets.MNIST(root="../DataSet/mnist",
train=True,
transform=transform,
download=False)
test_dataset = datasets.MNIST(root="../DataSet/mnist",
train=False,
transform=transform,
download=False)
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)
2. 構(gòu)建GoogleNet---構(gòu)造Inception單元
# 定義GoogleNet---構(gòu)造 Inception 單元---GoogleNet不改變圖片的尺寸大小即 w 和 h 不變栏饮，只改變其 channel 大小
class GoogleNet(torch.nn.Module):
def __init__(self, input_channels):
super(GoogleNet, self).__init__()
# 第一個(gè)分支
self.branch_pool1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
# 第二個(gè)分支
self.branch_pool2_1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
self.branch_pool2_2 = torch.nn.Conv2d(16, 24, kernel_size=5, padding=2)
# 第三個(gè)分支
# ---padding = (k-1)/2 k為卷積核大小, 即可保證卷積后圖片大小不變
# ---padding作用: 保證卷積后的圖片大小與原圖片一致即作用于 w 和 h
# ---同時(shí)用0填充,保證足夠多的信息量也不存在噪音問(wèn)題
self.branch_pool3_1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
self.branch_pool3_2 = torch.nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch_pool3_3 = torch.nn.Conv2d(24, 24, kernel_size=3, padding=1)
# 第四個(gè)分支
self.branch_pool4 = torch.nn.Conv2d(input_channels, 24, kernel_size=1)
# 不管 input_channels 是多少，輸出的 channels 是24*3+16=88
def forward(self, x):
branch1 = self.branch_pool1(x) # torch.Size([64, 16, 28, 28])
branch2 = self.branch_pool2_2(self.branch_pool2_1(x)) # torch.Size([64, 24, 28, 28])
branch3 = self.branch_pool3_3(self.branch_pool3_2(self.branch_pool3_1(x))) # torch.Size([64, 24, 28, 28])
branch4 = self.branch_pool4(F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)) # torch.Size([64, 24, 28, 28])
outputs = [branch1, branch2, branch3, branch4]
# GoogleNet 要求每個(gè)分支輸出的圖片通道數(shù)可以不一樣, 但其他維度的尺寸必須一樣,這樣才可以保證能做 cat 連接
# 即(batch, channel_branch, width, height)----->(batch, channel_branch, width, height)
# (batch, channels, width, height): 沿著 channel 通道的方向?qū)⑦@些張量連接起來(lái)
return torch.cat(outputs, dim=1)
3.采用GoogleNet的神經(jīng)網(wǎng)絡(luò)來(lái)構(gòu)建模型
# 2.構(gòu)建網(wǎng)絡(luò)模型---模型是針對(duì)批次樣本的處理情況
class Module(torch.nn.Module):
def __init__(self):
super(Module, self).__init__()
self.conv1 = torch.nn.Conv2d(1, 10, kernel_size=5, bias=False)
self.googleNet1 = GoogleNet(input_channels=10)
self.googleNet2 = GoogleNet(input_channels=20)
self.conv2 = torch.nn.Conv2d(88, 20, kernel_size=5, bias=False)
# 下采樣并不改變 channel 數(shù)量,只改變圖片大小
self.maxPooling = torch.nn.MaxPool2d(2)
self.fc = torch.nn.Linear(1408, 10)
# 卷積---池化---激活函數(shù)---GoogleNet---數(shù)據(jù)扁平化處理---全連接層
def forward(self, x):
size = x.size(0) # torch.Size([64, 1, 28, 28])
x = F.relu(self.maxPooling(self.conv1(x))) # torch.Size([64, 10, 12, 12])
x = self.googleNet1(x) # torch.Size([64, 88, 12, 12])
x = F.relu(self.maxPooling(self.conv2(x))) # torch.Size([64, 20, 4, 4])
x = self.googleNet2(x) # torch.Size([64, 88, 4, 4])
# 數(shù)據(jù)扁平化處理,為接下來(lái)的全連接測(cè)做準(zhǔn)備
# Flatten data from (64, 88, 4, 4) to (64,1408)
x = x.view(size, -1)
x = self.fc(x)
# 全連接層之后不需要跟激活函數(shù),因?yàn)榧せ詈瘮?shù) softmax 的作用包含在 CrossEntropyLoss 中
# softmax 函數(shù)的作用包含在 CrossEntropyLoss 中
return x
4. 構(gòu)建損失函數(shù)和優(yōu)化器
損失函數(shù)采用CrossEntropyLoss
優(yōu)化器采用 SGD 隨機(jī)梯度優(yōu)化算法
# 構(gòu)造損失器和優(yōu)化器
# softmax 函數(shù)的作用包含在 CrossEntropyLoss 中,交叉熵算法
criterion = torch.nn.CrossEntropyLoss()
opt = optim.SGD(params=model.parameters(), lr=0.01, momentum=0.5)
# 動(dòng)態(tài)更新學(xué)習(xí)率------每隔step_size : lr = lr * gamma
schedule = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.5, last_epoch=-1)
5.完整代碼
# -*- codeing = utf-8 -*-
# @Time : 2022/4/12 8:57
# @Software : PyCharm
# 超參數(shù): 訓(xùn)練之前設(shè)置的參數(shù)
# 模型參數(shù): 訓(xùn)練過(guò)程中得到的參數(shù)
# Average Pooling: 均值池化---
# network in network:
# 1*1 的卷積核: 單個(gè)1*1的卷積后的圖片大小不變即：c*w*h---------->1*w*h
# 若需要輸出16個(gè)通道的圖片磷仰，則只需要將輸出通道設(shè)置為16,pytorch自動(dòng)構(gòu)建16個(gè)通道,1*1的卷積核
# 1*1 的卷積不改變圖片尺寸,即c1*w*h---------->c2*w*h
# 1*1 的卷積可以有效降低通道數(shù)量,大大減少網(wǎng)絡(luò)模型的浮點(diǎn)數(shù)運(yùn)算量
import torch
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
# Maxpooling: 最大池化,尋找每個(gè)空間的最大值然后組成一個(gè)新的圖像
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(), # 轉(zhuǎn)化成Tensor張量
transforms.Normalize((0.1307,), (0.3081,)) # 歸一化處理,將其(0-255)映射到(0-1)
])
# 1.準(zhǔn)備數(shù)據(jù)集
train_dataset = datasets.MNIST(root="../DataSet/mnist",
train=True,
transform=transform,
download=False)
test_dataset = datasets.MNIST(root="../DataSet/mnist",
train=False,
transform=transform,
download=False)
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)
# 定義GoogleNet---構(gòu)造 Inception 單元---GoogleNet不改變圖片的尺寸大小即 w 和 h 不變袍嬉，只改變其 channel 大小
class GoogleNet(torch.nn.Module):
def __init__(self, input_channels):
super(GoogleNet, self).__init__()
# 第一個(gè)分支
self.branch_pool1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
# 第二個(gè)分支
self.branch_pool2_1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
self.branch_pool2_2 = torch.nn.Conv2d(16, 24, kernel_size=5, padding=2)
# 第三個(gè)分支
# ---padding = (k-1)/2 k為卷積核大小, 即可保證卷積后圖片大小不變
# ---padding作用: 保證卷積后的圖片大小與原圖片一致即作用于 w 和 h
# ---同時(shí)用0填充,保證足夠多的信息量也不存在噪音問(wèn)題
self.branch_pool3_1 = torch.nn.Conv2d(input_channels, 16, kernel_size=1)
self.branch_pool3_2 = torch.nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch_pool3_3 = torch.nn.Conv2d(24, 24, kernel_size=3, padding=1)
# 第四個(gè)分支
self.branch_pool4 = torch.nn.Conv2d(input_channels, 24, kernel_size=1)
# 不管 input_channels 是多少，輸出的 channels 是24*3+16=88
def forward(self, x):
branch1 = self.branch_pool1(x) # torch.Size([64, 16, 28, 28])
branch2 = self.branch_pool2_2(self.branch_pool2_1(x)) # torch.Size([64, 24, 28, 28])
branch3 = self.branch_pool3_3(self.branch_pool3_2(self.branch_pool3_1(x))) # torch.Size([64, 24, 28, 28])
branch4 = self.branch_pool4(F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)) # torch.Size([64, 24, 28, 28])
outputs = [branch1, branch2, branch3, branch4]
# GoogleNet 要求每個(gè)分支輸出的圖片通道數(shù)可以不一樣, 但其他維度的尺寸必須一樣,這樣才可以保證能做 cat 連接
# 即(batch, channel_branch, width, height)----->(batch, channel_branch, width, height)
# (batch, channels, width, height): 沿著 channel 通道的方向?qū)⑦@些張量連接起來(lái)
return torch.cat(outputs, dim=1)
# 2.構(gòu)建網(wǎng)絡(luò)模型---模型是針對(duì)批次樣本的處理情況
class Module(torch.nn.Module):
def __init__(self):
super(Module, self).__init__()
self.conv1 = torch.nn.Conv2d(1, 10, kernel_size=5, bias=False)
self.googleNet1 = GoogleNet(input_channels=10)
self.googleNet2 = GoogleNet(input_channels=20)
self.conv2 = torch.nn.Conv2d(88, 20, kernel_size=5, bias=False)
# 下采樣并不改變 channel 數(shù)量,只改變圖片大小
self.maxPooling = torch.nn.MaxPool2d(2)
self.fc = torch.nn.Linear(1408, 10)
# 卷積---池化---激活函數(shù)---GoogleNet---數(shù)據(jù)扁平化處理---全連接層
def forward(self, x):
size = x.size(0) # torch.Size([64, 1, 28, 28])
x = F.relu(self.maxPooling(self.conv1(x))) # torch.Size([64, 10, 12, 12])
x = self.googleNet1(x) # torch.Size([64, 88, 12, 12])
x = F.relu(self.maxPooling(self.conv2(x))) # torch.Size([64, 20, 4, 4])
x = self.googleNet2(x) # torch.Size([64, 88, 4, 4])
# 數(shù)據(jù)扁平化處理,為接下來(lái)的全連接測(cè)做準(zhǔn)備
# Flatten data from (64, 88, 4, 4) to (64,1408)
x = x.view(size, -1)
x = self.fc(x)
# 全連接層之后不需要跟激活函數(shù),因?yàn)榧せ詈瘮?shù) softmax 的作用包含在 CrossEntropyLoss 中
# softmax 函數(shù)的作用包含在 CrossEntropyLoss 中
return x
model = Module().to(device)
# 3.構(gòu)造損失器和優(yōu)化器
# softmax 函數(shù)的作用包含在 CrossEntropyLoss 中,交叉熵算法
criterion = torch.nn.CrossEntropyLoss()
opt = optim.SGD(params=model.parameters(), lr=0.01, momentum=0.5)
# 動(dòng)態(tài)更新學(xué)習(xí)率------每隔step_size : lr = lr * gamma
schedule = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.5, last_epoch=-1)
# 4.訓(xùn)練數(shù)據(jù)集
def train():
running_loss = 0
for batch_idx, (inputs, target) in enumerate(train_loader, 0):
inputs, target = inputs.to(device), target.to(device)
opt.zero_grad()
y_pred_data = model(inputs)
loss = criterion(y_pred_data, target)
loss.backward()
opt.step()
running_loss += loss.item()
if batch_idx % 300 == 299:
print("[%5d, %5d] loss: %.5f" % (epoch + 1, batch_idx + 1, running_loss / 300))
running_loss == 0.0
# 5.測(cè)試數(shù)據(jù)集
def verify():
correct = 0
total = 0
with torch.no_grad(): # 該語(yǔ)句下的所有tensor在進(jìn)行反向傳播時(shí),不會(huì)被計(jì)算梯度
for (images, labels) in test_loader:
images, labels = images.to(device), labels.to(device)
# 數(shù)據(jù)進(jìn)入模型進(jìn)行計(jì)算
outputs = model(images)
# 沿著維度為1的方向(行方向) 尋找每行最大元素的值與其下標(biāo)
_, predicted = torch.max(outputs.data, dim=1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print("==============================")
print("Accuracy on test set: %d%%" % (100 * correct / total))
print("==============================")
if __name__ == '__main__':
for epoch in range(15):
train()
verify()
# GoogleNet: 分支卷積--->cat(dim=1:channels)---全連接
# 使用 卷積 + GoogleNet + 全連接 的神經(jīng)網(wǎng)絡(luò)的準(zhǔn)確率在 99% 左右, 同時(shí)減少了參數(shù)量和計(jì)算量
6.結(jié)果展示
result.png
