cifar-10基础优化二pytorch，测试集准确率88.2

须知：

同前文

代码变化部分如下图，其余部分与上文一同CIFAR-10基础优化一（加入标准化和激活函数）_百炼成丹的博客-CSDN博客

优化思路：

再加入多层卷积层

卷积层层间和全连接层层间加入dropout

优化网络顺序结构

图像增广

网络结构：

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.model1 = nn.Sequential(
            nn.Conv2d(3, 32, 5, padding=2), nn.BatchNorm2d(32),
            nn.Dropout2d(0.25), nn.ReLU(),
            nn.Conv2d(32, 32, 5, padding=2), nn.BatchNorm2d(32),
            nn.MaxPool2d(2))  # nn.ReLU(),

        self.model2 = nn.Sequential(
            nn.Conv2d(32, 64, 5, padding=2), nn.BatchNorm2d(64),
            nn.Dropout2d(0.25), nn.ReLU(),
            nn.Conv2d(32, 64, 5, padding=2), nn.BatchNorm2d(64),
            nn.MaxPool2d(2))  # nn.ReLU(),

        self.model3 = nn.Sequential(
            nn.Conv2d(32, 128, 5, padding=2), nn.BatchNorm2d(128),
            nn.Dropout2d(0.3), nn.ReLU(),
            nn.Conv2d(32, 128, 5, padding=2), nn.BatchNorm2d(128),
            nn.MaxPool2d(2))  # nn.ReLU(),

        self.model4 = nn.Sequential(
            nn.Conv2d(32, 256, 5, padding=2), nn.BatchNorm2d(256),
            nn.Dropout2d(0.3), nn.ReLU(),
            nn.Conv2d(32, 256, 5, padding=2), nn.BatchNorm2d(256),
            nn.MaxPool2d(2))  # nn.ReLU(),

        self.model5 = nn.Sequential(
            nn.Conv2d(256, 512, 5, padding=2), nn.BatchNorm2d(512),
            nn.Dropout2d(0.35), nn.ReLU(),
            nn.Conv2d(512, 512, 5, padding=2), nn.BatchNorm2d(512),
            nn.MaxPool2d(2))  # nn.ReLU(),
        
        self.linear1 = nn.Sequential(
            nn.Flatten(),
            nn.Linear(512, 128), nn.BatchNorm1d(128), nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(128, 10))

图像增广：torchvision.transforms.ColorJitter(0.5), torchvision.transforms.RandomHorizontalFlip()

train_data = torchvision.datasets.CIFAR10("../dataset", train=True, transform=torchvision.transforms.Compose(
    [torchvision.transforms.ColorJitter(0.5), torchvision.transforms.RandomHorizontalFlip(),
     torchvision.transforms.ToTensor()]))

效果：

 测试集准确率峰值88.2%

 约64epoch后无增幅

表现效果进一步提高

但时间不理想，约40分钟上下

下一步，加入残差网络