街景字符识别task4学习笔记

在baseline基础上尝试过多种可能提高模型预测精度的思路，其中只有两种获得了一定程度的提升：（1）增加数据扩增多样性，降低过拟合；（2）将全连接层改为卷积层。

1 增加数据扩增多样性

train_loader = torch.utils.data.DataLoader(
    SVHNDataset(train_path, train_label,
                transforms.Compose([
                    transforms.Resize((64, 128)),#调整分辨率(h,w)
                    transforms.RandomCrop((60, 120)),#(h,w)
                    transforms.ColorJitter(0.3, 0.3, 0.2),
                    transforms.RandomRotation(5),#在(-5,5)之间随机旋转
                    transforms.RandomAffine(30,translate=(0, 0.2)),#仿射变换
                    transforms.RandomGrayscale(0.5),#转灰度图
                    transforms.RandomPerspective(distortion_scale=0.5, p=0.5, interpolation=3),#透视变换
                    transforms.ToTensor(),
                    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])#(mean,std)
    ])), 
    batch_size=40, 
    shuffle=True, 
    num_workers=10,
)

baseline算法在10个epoch后验证集精度为0.548，增加数据扩增方法后验证集精度为0.557，提高了一个百分点，训练结果如下。

2 将全连接层改为卷积层

class SVHN_Model1(nn.Module):
    def __init__(self):
        super(SVHN_Model1, self).__init__()
                
        model_conv = models.resnet18(pretrained=True)
        model_conv.avgpool = nn.AdaptiveAvgPool2d(1)
        model_conv = nn.Sequential(*list(model_conv.children())[:-1])
        self.cnn = model_conv
        
        self.cv1 = nn.Conv2d(512,11,kernel_size=1)
        self.cv2 = nn.Conv2d(512,11,kernel_size=1)
        self.cv3 = nn.Conv2d(512,11,kernel_size=1)
        self.cv4 = nn.Conv2d(512,11,kernel_size=1)
        self.cv5 = nn.Conv2d(512,11,kernel_size=1)
        
        '''
        self.fc1 = nn.Linear(512, 11)
        self.fc2 = nn.Linear(512, 11)
        self.fc3 = nn.Linear(512, 11)
        self.fc4 = nn.Linear(512, 11)
        self.fc5 = nn.Linear(512, 11)
        '''
    def forward(self, img):        
        feat = self.cnn(img)
        # print(feat.shape)
        c1 = torch.squeeze(self.cv1(feat))
        c2 = torch.squeeze(self.cv2(feat))
        c3 = torch.squeeze(self.cv3(feat))
        c4 = torch.squeeze(self.cv4(feat))
        c5 = torch.squeeze(self.cv5(feat))
        
        '''
        feat = feat.view(feat.shape[0], -1)
        c1 = self.fc1(feat)
        c2 = self.fc2(feat)
        c3 = self.fc3(feat)
        c4 = self.fc4(feat)
        c5 = self.fc5(feat)
        '''
        return c1, c2, c3, c4, c5

10个epoch后验证集精度又提高了一个百分点，为0.567，训练结果如下：