CIFAR-10數(shù)據(jù)集介紹
CIFAR-10數(shù)據(jù)集由10個(gè)類的60000個(gè)32x32彩色圖像組成帝蒿，每個(gè)類有6000個(gè)圖像。有50000個(gè)訓(xùn)練圖像和10000個(gè)測試圖像曙咽。
數(shù)據(jù)集分為五個(gè)訓(xùn)練批次和一個(gè)測試批次省店，每個(gè)批次有10000個(gè)圖像。測試批次包含來自每個(gè)類別的恰好1000個(gè)隨機(jī)選擇的圖像宙拉。訓(xùn)練批次以隨機(jī)順序包含剩余圖像宾尚，但一些訓(xùn)練批次可能包含來自一個(gè)類別的圖像比另一個(gè)更多⌒怀海總體來說煌贴，五個(gè)訓(xùn)練集之和包含來自每個(gè)類的正好5000張圖像。
以下是數(shù)據(jù)集中的類锥忿，以及來自每個(gè)類的10個(gè)隨機(jī)圖像：
CIFAR10數(shù)據(jù)集
下載地址：http://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
切入正題
這次實(shí)踐主要參考PyTorch官方的教程（https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py）牛郑，在此基礎(chǔ)上進(jìn)行一些修改，由于主要目的是了解PyTorch的編程方法敬鬓，所以在數(shù)據(jù)集那里并沒有從訓(xùn)練集中切分驗(yàn)證集出來淹朋，在訓(xùn)練時(shí)僅觀察了loss的變化，最后使用測試集觀察準(zhǔn)確率钉答。
關(guān)于數(shù)據(jù)集
CIFAR10的數(shù)據(jù)集可以通過torchvision進(jìn)行下載础芍，但是下載速度太慢，建議使用迅雷下載
開搞
導(dǎo)入基礎(chǔ)包
import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
# 設(shè)置一些參數(shù)
EPOCHS = 20
BATCH_SIZE = 512
創(chuàng)建數(shù)據(jù)集
# 創(chuàng)建一個(gè)轉(zhuǎn)換器数尿，將torchvision數(shù)據(jù)集的輸出范圍[0,1]轉(zhuǎn)換為歸一化范圍的張量[-1,1]者甲。
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
# 創(chuàng)建訓(xùn)練集
# root -- 數(shù)據(jù)存放的目錄
# train -- 明確是否是訓(xùn)練集
# download -- 是否需要下載
# transform -- 轉(zhuǎn)換器，將數(shù)據(jù)集進(jìn)行轉(zhuǎn)換
trainset = torchvision.datasets.CIFAR10(
root='./data',
train=True,
download=True,
transform=transform
)
# 創(chuàng)建測試集
testset = torchvision.datasets.CIFAR10(
root='./data',
train=False,
download=True,
transform=transform
)
創(chuàng)建數(shù)據(jù)加載器
# 創(chuàng)建訓(xùn)練/測試加載器砌创，
# trainset/testset -- 數(shù)據(jù)集
# batch_size -- 不解釋
# shuffle -- 是否打亂順序
trainloader = torch.utils.data.DataLoader(
trainset, batch_size=BATCH_SIZE, shuffle=True)
testloader = torch.utils.data.DataLoader(
testset, batch_size=BATCH_SIZE, shuffle=True)
# 類別標(biāo)簽
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
創(chuàng)建網(wǎng)絡(luò)
我這里定義了兩個(gè)CNN網(wǎng)絡(luò)虏缸，分別保存在CNN_1.py和CNN_2.py文件中
CNN_1
# CNN_1.py
import torch
# 學(xué)習(xí)率
LR = 0.005
class Net(torch.nn.Module):
"""Some Information about Net"""
def __init__(self):
super(Net, self).__init__()
self.conv1 = torch.nn.Sequential(
torch.nn.Conv2d(3, 16, 3, padding=1), # 3*32*32 -> 16*32*32
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 16*32*32 -> 16*16*16
)
self.conv2 = torch.nn.Sequential(
torch.nn.Conv2d(16, 32, 3, padding=1), # 16*16*16 -> 32*16*16
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 32*16*16 -> 32*8*8
)
self.conv3 = torch.nn.Sequential(
torch.nn.Conv2d(32, 64, 3, padding=1), # 32*8*8 -> 64*8*8
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 64*8*8 -> 64*4*4
)
self.fc1 = torch.nn.Sequential(
torch.nn.Linear(64*4*4, 32),
torch.nn.ReLU(),
# torch.nn.Dropout()
)
self.fc2 = torch.nn.Linear(32, 10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = x.view(-1, 64*4*4)
x = self.fc1(x)
x = self.fc2(x)
return x
net = Net()
net.cuda()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=LR)
網(wǎng)絡(luò)結(jié)構(gòu)如下：
Net(
(conv1): Sequential(
(0): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv2): Sequential(
(0): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv3): Sequential(
(0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(fc1): Sequential(
(0): Linear(in_features=1024, out_features=32, bias=True)
(1): ReLU()
)
(fc2): Linear(in_features=32, out_features=10, bias=True)
)
網(wǎng)絡(luò)創(chuàng)建了3個(gè)卷積層鲫懒，1個(gè)全連接層
CNN_2
import torch
LR = 0.005
class Net(torch.nn.Module):
"""Some Information about CNN"""
def __init__(self):
super(Net, self).__init__()
self.conv1 = torch.nn.Sequential(
torch.nn.Conv2d(3, 16, 3, padding=1), # 3*32*32 -> 16*32*32
torch.nn.ReLU(),
)
self.conv2 = torch.nn.Sequential(
torch.nn.Conv2d(16, 32, 3, padding=1), # 16*32*32 -> 32*32*32
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 32*32*32-> 32*16*16
)
self.conv3 = torch.nn.Sequential(
torch.nn.Conv2d(32, 64, 3, padding=1), # 32*16*16 -> 64*16*16
torch.nn.ReLU(),
)
self.conv4 = torch.nn.Sequential(
torch.nn.Conv2d(64, 128, 3, padding=1), # 64*16*16 -> 128*16*16
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 128*16*16 -> 128*8*8
)
self.conv5 = torch.nn.Sequential(
torch.nn.Conv2d(128, 256, 3, padding=1), # 128*8*8 -> 256*8*8
torch.nn.ReLU(),
torch.nn.MaxPool2d(2, 2) # 256*8*8 -> 256*4*4
)
self.gap = torch.nn.AvgPool2d(4,4)
self.fc = torch.nn.Linear(256, 10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = self.conv4(x)
x = self.conv5(x)
x = self.gap(x)
x = x.view(-1, 256)
x = self.fc(x)
return x
net = Net()
net.cuda()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=LR)
打印網(wǎng)絡(luò)如下：
Net(
(conv1): Sequential(
(0): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
)
(conv2): Sequential(
(0): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv3): Sequential(
(0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
)
(conv4): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv5): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(gap): AvgPool2d(kernel_size=4, stride=4, padding=0)
(fc): Linear(in_features=256, out_features=10, bias=True)
)
創(chuàng)建了5個(gè)卷積層，最后使用GAP連接輸出層
定義訓(xùn)練函數(shù)和測試函數(shù)
為了方便對不同的網(wǎng)絡(luò)進(jìn)行訓(xùn)練和測試刽辙，因此窥岩，定義一個(gè)通用的訓(xùn)練函數(shù)和測試函數(shù)，方便使用
定義訓(xùn)練函數(shù)
def train(model, criterion, optimizer, trainloader, epochs=5, log_interval=50):
print('----- Train Start -----')
for epoch in range(epochs):
running_loss = 0.0
for step, (batch_x, batch_y) in enumerate(trainloader):
batch_x, batch_y = batch_x.cuda(), batch_y.cuda()
output = model(batch_x)
optimizer.zero_grad()
loss = criterion(output, batch_y)
loss.backward()
optimizer.step()
running_loss += loss.item()
if step % log_interval == (log_interval-1):
print('[%d, %5d] loss: %.4f' %
(epoch + 1, step + 1, running_loss / log_interval))
running_loss = 0.0
print('----- Train Finished -----')
定義測試函數(shù)
def test(model, testloader):
print('------ Test Start -----')
correct = 0
total = 0
with torch.no_grad():
for test_x, test_y in testloader:
images, labels = test_x.cuda(), test_y.cuda()
output = model(images)
_, predicted = torch.max(output.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
accuracy = 100 * correct / total
print('Accuracy of the network is: %.4f %%' % accuracy)
return accuracy
在測試集上運(yùn)行網(wǎng)絡(luò)
測試CNN_1網(wǎng)絡(luò)
train(CNN_1.net, CNN_1.criterion, CNN_1.optimizer, trainloader, epochs=EPOCHS)
test(CNN_1.net, testloader)
測試CNN_2網(wǎng)絡(luò)
train(CNN_2.net, CNN_2.criterion, CNN_2.optimizer, trainloader, epochs=EPOCHS)
test(CNN_2.net, testloader)
訓(xùn)練結(jié)果
CNN_1在10代訓(xùn)練后宰缤，在測試集準(zhǔn)確率上能夠達(dá)到71.1100 %
CNN_2在10代訓(xùn)練后颂翼，在測試集準(zhǔn)確率上能夠達(dá)到75.6700 %
代碼
鏈接：https://pan.baidu.com/s/1rOmiE35rQnszmYyyN6h16Q
提取碼：zqf2
