[Pytorch] SPP-Net 3D版本原理與實現

代碼參考2d的 spp-net code Github

1.SPP 插入位置

在分類之前，FC要求輸入特定尺寸的feature。當輸入中包含不同size的圖像時，傳統的fc就不可以了。此時kaiming提出了SPP-Net，解決了這個問題。

2. SPP module 細節

輸入任意大小的feature，使用maxpooling/average pooling等方法得到4x4xc,2x2xc, 1x1xc的三個固定大小feature。

3.代碼實現

class SpatialPyramidPool3D(nn.Module):
    """
    Args:
        out_side (tuple): Length of side in the pooling results of each pyramid layer.

    Inputs:
        - `input`: the input Tensor to invert ([batch, channel, width, height])
    """

    def __init__(self, out_side):
        super(SpatialPyramidPool3D, self).__init__()
        self.out_side = out_side

    def forward(self, x):
        out = None
        for n in self.out_side:
            d_r, w_r, h_r = map(lambda s: math.ceil(s / n), x.size()[2:])  # Receptive Field Size
            s_d, s_w, s_h = map(lambda s: math.floor(s / n), x.size()[2:])  # Stride
            max_pool = nn.MaxPool3d(kernel_size=(d_r, w_r, h_r), stride=(s_d, s_w, s_h))
            y = max_pool(x)
            if out is None:
                out = y.view(y.size()[0], -1)
            else:
                out = torch.cat((out, y.view(y.size()[0], -1)), 1)
        return out

用的時候這麼用，如果feature本身大小小於4*4，那肯定不行，4去掉就行。

self.spp = SpatialPyramidPool3D(out_side=(1,2,4))