CIFAR-100数据集卷积神经网络训练

1. CIFAR-10数据集介绍

2. 问题说明

3. 模型训练过程

4. 结果可视化

1. CIFAR-100数据集介绍

这个数据集就像CIFAR-10，除了它有100个类，每个类包含600个图像。，每类各有500个训练图像和100个测试图像。CIFAR-100中的100个类被分成20个超类。每个图像都带有一个“精细”标签（它所属的类）和一个“粗糙”标签（它所属的超类）以下是CIFAR-100中的类别列表：

Superclass	Classes
aquatic	mammals beaver, dolphin, otter, seal, whale
fish	aquarium fish, flatfish, ray, shark, trout
flowers	orchids, poppies, roses, sunflowers, tulips
food	containers bottles, bowls, cans, cups, plates
fruit and vegetables	apples, mushrooms, oranges, pears, sweet peppers
household electrical devices	clock, computer keyboard, lamp, telephone, television
household	furniture bed, chair, couch, table, wardrobe
insects	bee, beetle, butterfly, caterpillar, cockroach
large carnivores	bear, leopard, lion, tiger, wolf
large man-made outdoor things	bridge, castle, house, road, skyscraper
large natural outdoor scenes	cloud, forest, mountain, plain, sea
large omnivores and herbivores	camel, cattle, chimpanzee, elephant, kangaroo
medium-sized mammals	fox, porcupine, possum, raccoon, skunk
non-insect invertebrates	crab, lobster, snail, spider, worm
people	baby, boy, girl, man, woman
reptiles	crocodile, dinosaur, lizard, snake, turtle
small mammals	hamster, mouse, rabbit, shrew, squirrel
trees	maple, oak, palm, pine, willow
vehicles 1	bicycle, bus, motorcycle, pickup truck, train
vehicles 2	lawn-mower, rocket, streetcar, tank, tractor

CIFAR-100标签

参考

2. 问题说明

自由选择算法完成 cifar-100 数据集解析分类
要求准确率至少大于60%

3. 模型训练过程

1.初步准备

采用ResNet模型

两种残差网络块定义

'''
不带瓶颈段的残差块
'''
class BasicBlock(nn.Module):def __init__(self, in_channels, out_channels, stride=1):super(BasicBlock, self).__init__()self.network = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, stride=stride),nn.BatchNorm2d(out_channels),nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1, stride=1),nn.MaxPool2d(3, stride=1, padding=1))self.downSample = nn.Sequential()if in_channels != out_channels:self.downSample = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride),nn.BatchNorm2d(out_channels))def forward(self, x):out = self.network(x) + self.downSample(x)return out

'''
带瓶颈段的残差块
in channel = [inChannel1,inChannel2,inChannel3]
in channel = [outChannel1,outChannel2,outChannel3]
stride = 1 长宽不变
stride = 2 长宽减半
'''
class BasicBlock(nn.Module):def __init__(self, in_channels, out_channels, stride):super(BasicBlock, self).__init__()self.network = nn.Sequential(nn.Conv2d(in_channels, out_channels[0], kernel_size=1, padding=0, stride=1),nn.BatchNorm2d(out_channels[0]),nn.ReLU(),nn.Conv2d(out_channels[0], out_channels[1], kernel_size=3, padding=1, stride=stride),nn.BatchNorm2d(out_channels[1]),nn.ReLU(),nn.Conv2d(out_channels[1], out_channels[2], kernel_size=1, padding=0, stride=1),nn.BatchNorm2d(out_channels[2]),)self.downSample = nn.Sequential(nn.Conv2d(in_channels, out_channels[2], kernel_size=1,  padding=0, stride=stride),nn.BatchNorm2d(out_channels[2]))

训练，测试函数定义如下

def train():net.train()acc = 0.0sum = 0.0loss_sum = 0for batch, (data, target) in enumerate(trainLoader):data, target = data.to(device), target.to(device)net.optimizer.zero_grad()output = net(data)loss = net.lossFunc(output, target)loss.backward()net.optimizer.step()acc += torch.sum(torch.argmax(output, dim=1) == target).item()sum += len(target)loss_sum += loss.item()writer.add_scalar('Cifar100_model_log/trainAccuracy', 100 * acc / sum, epoch + 1)writer.add_scalar('Cifar100_model_log/trainLoss', loss_sum / (batch + 1), epoch + 1)print('train accuracy: %.2f%%, loss: %.4f' % (100 * acc / sum, loss_sum / (batch + 1)))def test():net.eval()acc = 0.0sum = 0.0loss_sum = 0step = 0for batch, (data, target) in enumerate(testLoader):initData = datadata = testTransform2(data)data, target = data.to(device), target.to(device)output = net(data)
'''
用于通过tensorboard可视化训练结果，使用时要注意将epoch设置为1for i in range(len(output)):writer.add_image(fineLabelList[torch.argmax(output, dim=1)[i]], initData[i], step)step = step + 1
'''loss = net.lossFunc(output, target)acc += torch.sum(torch.argmax(output, dim=1) == target).item()sum += len(target)loss_sum += loss.item()writer.add_scalar('Cifar100_model_log/testAccuracy', 100 * acc / sum, epoch + 1)writer.add_scalar('Cifar100_model_log/trainLoss', loss_sum / (batch + 1), epoch + 1)print('test accuracy: %.2f%%, loss: %.4f' % (100 * acc / sum, loss_sum / (batch + 1)))

准备数据

'''
采用GPU加速训练，没有则用cpu
'''
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
trainTransform = transforms.Compose([transforms.ToTensor(),transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])'''
数据集扩充采用
trainTransform = transforms.Compose([transforms.ToTensor(),transforms.RandomCrop(32, padding=4),transforms.RandomHorizontalFlip(),transforms.ColorJitter(brightness=(0,1), contrast=(0,1), saturation=(0,1), hue=0),transforms.RandomVerticalFlip(p=0.5),transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
'''testTransform1 = transforms.Compose([transforms.ToTensor(),
])
testTransform2 = transforms.Compose([transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])trainDataset = datasets.cifar.CIFAR100(root='cifar100', train=True, transform=trainTransform, download=False)
testDataset = datasets.cifar.CIFAR100(root='cifar100', train=False, transform=testTransform1, download=False)trainLoader = DataLoader(trainDataset, batch_size=200, shuffle=True)
testLoader = DataLoader(testDataset, batch_size=200, shuffle=False)

2.模型训练及改进思路

训练过程中发现网络的过拟合问题情况严重，训练集正确率可达90%，测试集正确率卡在45%上下。

解决方案

（1）使用L1、L2正则化

（2）增加Dropout

（3）扩充数据集

（4）删除部分残差块，缩减网络

（5）迁移学习

经过测试，发现采用了缩减网络，扩充数据集，增加Dropout的网络可以使正确率从45%提升到了61.79%（迁移学习作用也比较明显，实验中的一种迁移学习方式在原训练基础上将正确率提升至63.13%）而相较于Dropout，L2正则化可以更好地防止过拟合，但是正确率上升缓慢且出现停滞。

最后网络定义如下，剩余残差块和卷积层可进行迁移训练

进行的部分迁移学习思路过程如下：

第一次迁移学习:
冻结block1~4
增加
block5：BasicBlock(128, 128)
block6：BasicBlock(128, 512, 2)
重设全连接层
正确率浮动变化不大第二次迁移学习：
第一次的基础上冻结所有层
补充cov1单层卷积2*2*512
（如果增加至三层则会下降严重）
正确率提升至58%第三次迁移学习
冻结block1~4
原版基础上重设全连接层单层
目的为了训练卷积层第四次迁移学习（较为成功）
冻结block1~4
在原版基础上增加
block5：BasicBlock(128, 1024)
block6：BasicBlock(1024, 128)
进行一次放大将正确率提升至63.13%
经测试若采用缩小，正确率提高幅度不大

class Net(nn.Module):def __init__(self):super(Net, self).__init__()self.block1 = BasicBlock(3, 16)self.block2 = BasicBlock(16, 64, 2)self.block3 = BasicBlock(64, 64)self.block4 = BasicBlock(64, 128, 2)self.block5 = nn.Sequential()self.block6 = nn.Sequential()self.block7 = nn.Sequential()self.block8 = nn.Sequential()self.block9 = nn.Sequential()self.cov1 = nn.Sequential()self.linear = nn.Sequential(nn.Flatten(),nn.Linear(8 * 8 * 128, 2048),nn.Dropout(0.1),nn.BatchNorm1d(2048),nn.Linear(2048, 1024),nn.Dropout(0.1),nn.ReLU(),nn.Linear(1024, 100))self.optimizer = torch.optim.SGD(self.parameters(), lr=0.08)self.lossFunc = torch.nn.CrossEntropyLoss()

训练结果（可以发现还是过拟合严重）

4. 结果可视化

受版面限制，部分展示如下

通过拖动发现，在bed标签下出现一只错误识别的兔子