resizer_model

Learning to Resize Images for Computer Vision Tasks
文章主题思想，使用网络进行学习，来调整输入图片的大小。

1. kaggel 中给出的实现方式

https://www.kaggle.com/c/seti-breakthrough-listen/discussion/246558；

class CNNWithResizer(nn.Module):
    def __init__(self, cfg, pretrained=False):
        super().__init__()
        self.cfg = cfg
        self.n = 16
        self.slope = .1
        self.r = 1
        self.cnn = timm.create_model(self.cfg.model_name, pretrained=pretrained, in_chans=1)

        if hasattr(self.cnn, "fc"):
            nb_ft = self.cnn.fc.in_features
            self.cnn.fc = nn.Identity()
        elif hasattr(self.cnn, "_fc"):
            nb_ft = self.cnn._fc.in_features
            self.cnn._fc = nn.Identity()
        elif hasattr(self.cnn, "classifier"):
            nb_ft = self.cnn.classifier.in_features
            self.cnn.classifier = nn.Identity()
        elif hasattr(self.cnn, "last_linear"):
            nb_ft = self.cnn.last_linear.in_features
            self.cnn.last_linear = nn.Identity()
        elif hasattr(self.cnn, "head"):
            nb_ft = self.cnn.head.in_features
            self.cnn.head = nn.Identity()

        self.block1 = nn.Sequential(
                nn.Conv2d(1, self.n, kernel_size=(7, 7), stride=(1,1), padding=(1, 1), bias=False),
                nn.LeakyReLU(negative_slope=self.slope),
                nn.Conv2d(self.n, self.n, kernel_size=(1, 1), stride=(1,1), padding=(1, 1), bias=False),
                nn.LeakyReLU(negative_slope=self.slope),
                nn.BatchNorm2d(self.n))
        self.block2 = nn.Sequential(
                nn.Conv2d(self.n, self.n, kernel_size=(3, 3), stride=(1,1), padding=(1, 1), bias=False),
                nn.BatchNorm2d(self.n),
                nn.LeakyReLU(negative_slope=self.slope),
                nn.Conv2d(self.n, self.n, kernel_size=(3, 3), stride=(1,1), padding=(1, 1), bias=False),
                nn.BatchNorm2d(self.n))
        self.block3 = nn.Sequential(
                nn.Conv2d(self.n, self.n, kernel_size=(3, 3), stride=(1,1), padding=(1, 1), bias=False),
                nn.BatchNorm2d(self.n))
        self.block4 = nn.Sequential(
                nn.Conv2d(self.n, 1, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3), bias=False))
        self.fc = nn.Linear(nb_ft, self.cfg.target_size)

    def forward(self, x):
        res1 = F.interpolate(x, size=(256, 256), mode='bilinear')
        x = self.block1(x)
        res2 = F.interpolate(x, size=(256, 256), mode='bilinear')
        x = self.block2(res2)
        x += res2
        if self.r > 1:
            for _ in range(self.r):
                res2 = x
                x = self.block2(x)
                x += res2
        x = self.block3(x)
        x += res2
        x = self.block4(x)
        x += res1
        x = self.cnn(x)
        x = self.fc(x)
        return x

1.1 github 中给出的实现方式

https://www.kushajveersingh.com/blog/data-augmentation-with-resizer-network-for-image-classification;

https://github.com/KushajveerSingh/resize_network_cv