从pytorch到keras的多尺度Cnn模型

Question

我在pytorch中有这个模型，它基本上是一个多尺度CNN模型，它返回最后为该图像连接的特征图。我不确定如何使用keras在tensorflow 2中编写相同的内容。任何指针表示赞赏。

class VGG16ScaledFeatures(object):

def __init__(self, last_layer=22):
    self.vgg16_features = torch.nn.ModuleList(
        list(models.vgg16(pretrained=True).features)[:last_layer]
    ).eval()

def __call__(self, org):
    x_ = torch.tensor([])
    with torch.no_grad():
        for s in range(3):
            x = F.max_pool2d(org, (2 ** s, 2 ** s))
            for i, f in enumerate(self.vgg16_features):
                x = f(x)
                if (
                    (s == 0 and i == 21)
                    or (s == 1 and i == 14)
                    or (s == 2 and i == 7)
                ):
                    x_ = torch.cat([x_, x], dim=1)
                    break

    x_ = (x_ - x_.mean(dim=(2, 3), keepdim=True)) / x_.std(dim=(2, 3), keepdim=True)

    return x_