CIFAR10数据集来源：torchvision.datasets.CIFAR10()
CIFAR10数据集是一个用于识别普适物体的小型数据集，一共包含10个类别的RGB彩色图片，图片尺寸大小为32x32，如图：
CIFAR10.png
相较于MNIST数据集，MNIST数据集是28x28的单通道灰度图，而CIFAR10数据集是32x32的RGB三通道彩色图，CIFAR10数据集更接近于真实世界的图片。
1. 数据集构建
每个像素点即每条数据中的值范围为0-255，有的数字过大不利于训练且难以收敛，故将其归一化到（0-1）之间
# 数据集处理
# transforms.RandomHorizontalFlip(p=0.5)---以0.5的概率对图片做水平横向翻转
transform_train = transforms.Compose([transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])
# transforms.ToTensor()---shape从(H,W,C)->(C,H,W), 每个像素点从(0-255)映射到(0-1):直接除以255
# transforms.Normalize---先将输入归一化到(0,1),像素点通过"(x-mean)/std",将每个元素分布到(-1,1)
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(std=(0.485, 0.456, 0.406), mean=(0.226, 0.224, 0.225))])
train_dataset = datasets.CIFAR10(root="../DataSet/cifar10", train=True, transform=transform_train,
download=True)
test_dataset = datasets.CIFAR10(root="../DataSet/cifar10", train=False, transform=transform,
download=True)
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)
2.构建DenseNet网络模型
1）构建 DenseNet-Block单元
class Bottleneck(nn.Module):
def __init__(self, input_channel, growth_rate):
super(Bottleneck, self).__init__()
self.bn1 = nn.BatchNorm2d(input_channel)
self.relu1 = nn.ReLU(inplace=True)
self.conv1 = nn.Conv2d(input_channel, 4 * growth_rate, kernel_size=1)
self.bn2 = nn.BatchNorm2d(4 * growth_rate)
self.relu2 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, padding=1)
def forward(self, x):
out = self.conv1(self.relu1(self.bn1(x)))
out = self.conv2(self.relu2(self.bn2(out)))
out = torch.cat([out, x], 1)
return out
2）Transition模块---连接两个DenseNet_Block
class Transition(nn.Module):
def __init__(self, input_channels, out_channels):
super(Transition, self).__init__()
self.bn = nn.BatchNorm2d(input_channels)
self.relu = nn.ReLU(inplace=True)
self.conv = nn.Conv2d(input_channels, out_channels, kernel_size=1)
def forward(self, x):
out = self.conv(self.relu(self.bn(x)))
out = F.avg_pool2d(out, 2)
return out
3. 构建损失函数和优化器
损失函数采用CrossEntropyLoss
优化器采用 SGD 随机梯度优化算法
# 构造损失函数和优化器
criterion = nn.CrossEntropyLoss()
opt = optim.SGD(model.parameters(), lr=0.01, momentum=0.8, weight_decay=5e-4)
# 动态更新学习率------每隔step_size : lr = lr * gamma
schedule = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.6, last_epoch=-1)
4.完整代码
import math
import torch.nn.functional as F
import torch
from torch import nn, optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms
from matplotlib import pyplot as plt
import time
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# transforms.RandomHorizontalFlip(p=0.5)---以0.5的概率对图片做水平横向翻转
transform_train = transforms.Compose([transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])
# transforms.ToTensor()---shape从(H,W,C)->(C,H,W), 每个像素点从(0-255)映射到(0-1):直接除以255
# transforms.Normalize---先将输入归一化到(0,1),像素点通过"(x-mean)/std",将每个元素分布到(-1,1)
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(std=(0.485, 0.456, 0.406), mean=(0.226, 0.224, 0.225))])
train_dataset = datasets.CIFAR10(root="../DataSet/cifar10", train=True, transform=transform_train,
download=True)
test_dataset = datasets.CIFAR10(root="../DataSet/cifar10", train=False, transform=transform,
download=True)
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)
class Bottleneck(nn.Module):
def __init__(self, input_channel, growth_rate):
super(Bottleneck, self).__init__()
self.bn1 = nn.BatchNorm2d(input_channel)
self.relu1 = nn.ReLU(inplace=True)
self.conv1 = nn.Conv2d(input_channel, 4 * growth_rate, kernel_size=1)
self.bn2 = nn.BatchNorm2d(4 * growth_rate)
self.relu2 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, padding=1)
def forward(self, x):
out = self.conv1(self.relu1(self.bn1(x)))
out = self.conv2(self.relu2(self.bn2(out)))
out = torch.cat([out, x], 1)
return out
class Transition(nn.Module):
def __init__(self, input_channels, out_channels):
super(Transition, self).__init__()
self.bn = nn.BatchNorm2d(input_channels)
self.relu = nn.ReLU(inplace=True)
self.conv = nn.Conv2d(input_channels, out_channels, kernel_size=1)
def forward(self, x):
out = self.conv(self.relu(self.bn(x)))
out = F.avg_pool2d(out, 2)
return out
class DenseNet(nn.Module):
def __init__(self, nblocks, growth_rate, reduction, num_classes):
super(DenseNet, self).__init__()
self.growth_rate = growth_rate
num_planes = 2 * growth_rate
self.basic_conv = nn.Sequential(
nn.Conv2d(3, 2 * growth_rate, kernel_size=7, stride=2, padding=3),
nn.BatchNorm2d(2 * growth_rate),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
)
self.dense1 = self._make_dense_layers(num_planes, nblocks[0])
num_planes += nblocks[0] * growth_rate
out_planes = int(math.floor(num_planes * reduction))
self.trans1 = Transition(num_planes, out_planes)
num_planes = out_planes
self.dense2 = self._make_dense_layers(num_planes, nblocks[1])
num_planes += nblocks[1] * growth_rate
out_planes = int(math.floor(num_planes * reduction))
self.trans2 = Transition(num_planes, out_planes)
num_planes = out_planes
self.dense3 = self._make_dense_layers(num_planes, nblocks[2])
num_planes += nblocks[2] * growth_rate
out_planes = int(math.floor(num_planes * reduction))
self.trans3 = Transition(num_planes, out_planes)
num_planes = out_planes
self.dense4 = self._make_dense_layers(num_planes, nblocks[3])
num_planes += nblocks[3] * growth_rate
self.AdaptiveAvgPool2d = nn.AdaptiveAvgPool2d(1)
# 全连接层
self.fc = nn.Sequential(
nn.Linear(num_planes, 256),
nn.ReLU(inplace=True),
# 使一半的神经元不起作用，防止参数量过大导致过拟合
nn.Dropout(0.5),
nn.Linear(256, 128),
nn.ReLU(inplace=True),
nn.Dropout(0.5),
nn.Linear(128, 10)
)
def _make_dense_layers(self, in_planes, nblock):
layers = []
for i in range(nblock):
layers.append(Bottleneck(in_planes, self.growth_rate))
in_planes += self.growth_rate
return nn.Sequential(*layers)
def forward(self, x):
out = self.basic_conv(x)
out = self.trans1(self.dense1(out))
out = self.trans2(self.dense2(out))
out = self.trans3(self.dense3(out))
out = self.dense4(out)
out = self.AdaptiveAvgPool2d(out)
out = out.view(out.size(0), -1)
out = self.fc(out)
return out
def DenseNet121():
return DenseNet([6, 12, 24, 16], growth_rate=32, reduction=0.5, num_classes=10)
def DenseNet169():
return DenseNet([6, 12, 32, 32], growth_rate=32, reduction=0.5, num_classes=10)
def DenseNet201():
return DenseNet([6, 12, 48, 32], growth_rate=32, reduction=0.5, num_classes=10)
def DenseNet265():
return DenseNet([6, 12, 64, 48], growth_rate=32, reduction=0.5, num_classes=10)
# 初始化模型
model = DenseNet121().to(device)
# 构造损失函数和优化器
criterion = nn.CrossEntropyLoss()
opt = optim.SGD(model.parameters(), lr=0.01, momentum=0.8, weight_decay=0.001)
# 动态更新学习率------每隔step_size : lr = lr * gamma
schedule = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.6, last_epoch=-1)
loss_list = []
# train
def train(epoch):
start = time.time()
for epoch in range(epoch):
running_loss = 0.0
for i, (inputs, labels) in enumerate(train_loader, 0):
inputs, labels = inputs.to(device), labels.to(device)
# 将数据送入模型训练
outputs = model(inputs)
# 计算损失
loss = criterion(outputs, labels).to(device)
# 重置梯度
opt.zero_grad()
# 计算梯度，反向传播
loss.backward()
# 根据反向传播的梯度值优化更新参数
opt.step()
# 100个batch的 loss 之和
running_loss += loss.item()
loss_list.append(loss.item())
# 每100个 batch 查看一下 平均loss
if (i + 1) % 100 == 0:
print('epoch = %d , batch = %d , loss = %.6f' % (epoch + 1, i + 1, running_loss / 100))
running_loss = 0.0
# 每一轮结束输出一下当前的学习率 lr
lr_1 = opt.param_groups[0]['lr']
print("learn_rate:%.15f" % lr_1)
schedule.step()
verify()
end = time.time()
# 计算并打印输出你的训练时间
print("time:{}".format(end - start))
# 训练过程可视化
plt.plot(loss_list)
plt.ylabel('loss')
plt.xlabel('Epoch')
plt.savefig('./DenseNet_train_img.png')
plt.show()
# Test
def verify():
model.eval()
correct = 0.0
total = 0
# 训练模式不需要反向传播更新梯度
with torch.no_grad():
print("===========================test===========================")
for inputs, labels in test_loader:
inputs, labels = inputs.to(device), labels.to(device)
outputs = model(inputs)
pred = outputs.argmax(dim=1) # 返回每一行中最大值元素索引
total += inputs.size(0)
correct += torch.eq(pred, labels).sum().item()
print("Accuracy of the network on the 10000 test images:%.2f %%" % (100 * correct / total))
print("==========================================================")
if __name__ == '__main__':
train(100)
verify()
# DenseNet: 所有卷积层全部使用使用3*3的卷积核, 两个3*3=一个5*5 同时可以减少参数量, 加深神经网络的深度
# 使用 DenseNet 神经网络训练 CIFAR10 数据集
