特征金字塔池化

1. 特征金字塔池化

如上图所示：

1. 将特征图的所有像素划分为$n\times n$个网格，对每个网格进行池化，池化层的核大小即为网格大小，宽度不符合时可以padding
2. 取不同的n值，重复1过程；
3. 将上述过程得到的所有结果经过flatten和concat,得到$C\times N$格式的特征图，可以直接用于全连接。

输出的结果只与$n$值和通道数量相关,而与输入Tensor的形状无关（当然不能太小，否则池化结果为0）

2. 实现

完整代码连接：古承风的gitee

以下是核心代码

def _spp_layer(self,x:torch.Tensor,mode='max',grid_nums:list=[16]):
        """
        output_num denote an grid's width
        
        steps:
        ---
        1. compute width for specific output_num, sqrt(num)
        2. compute pooling's kernel_size and stride
        3. pooling
        4. concat all the output
        """
        N,C,H,W = x.size()
        for i in range(len(grid_nums)):
            # step1
            
            h = ceil(H/(sqrt(grid_nums[i])))
            w = ceil(W/(sqrt(grid_nums[i])))
            
            h_pad = int(((h*sqrt(grid_nums[i])+1)-H)/2)
            w_pad = int(((w*sqrt(grid_nums[i])+1)-W)/2)
            # step2
            if mode == "max":
                pool = nn.MaxPool2d(kernel_size=(h,w),stride=(h,w),padding=(h_pad,w_pad)) 
            elif mode=='avg':
                pool = nn.AvgPool2d(kernel_size=(h,w),stride=(h,2),padding=(h_pad,w_pad))
            else:
                raise ValueError(f"{mode} mode type error ,expect 'max' and 'avg'")
            
            temp = pool(x) # to origin x , means pyramid pooling
            
            # if for fully connected , could use this concat method
            if i == 0:
                output = temp.view(N,-1)
            else:
                output = torch.concat((output,temp.view(N,-1)),-1)