pytorch中将3d卷积拆解成3x3卷积和3x1的卷积

       先要理清楚5D数据的形变问题，因为我们需要把数据降维，先来个简单的3D数据的降维再恢复原状：

    x = torch.arange(24).view(2,3,4)
    print(x)
    x = x.transpose(0,2)
    x = x.transpose(0,1)
    x = x.reshape(-1)
    print(x)
    x = x.reshape(4,3,2)
    x = x.transpose(2,0)
    x = x.transpose(1,2)
    x = x.reshape(2,3,4)
    print(x)

输出：

tensor([[[ 0,  1,  2,  3],
         [ 4,  5,  6,  7],
         [ 8,  9, 10, 11]],

        [[12, 13, 14, 15],
         [16, 17, 18, 19],
         [20, 21, 22, 23]]])
tensor([ 0, 12,  1, 13,  2, 14,  3, 15,  4, 16,  5, 17,  6, 18,  7, 19,  8, 20,
         9, 21, 10, 22, 11, 23])
tensor([[[ 0,  1,  2,  3],
         [ 4,  5,  6,  7],
         [ 8,  9, 10, 11]],

        [[12, 13, 14, 15],
         [16, 17, 18, 19],
         [20, 21, 22, 23]]])

再来个5D数据的降维和恢复原状：

    x = torch.arange(720).view(5,6,2,3,4)
    print(x)
    x = x.transpose(2,4)
    x = x.transpose(2,3)
    x = x.reshape(5,6,24)
    print(x)
    x = x.reshape(5,6,4,3,2)
    x = x.transpose(3,4)
    x = x.transpose(2,3)
    x = x.reshape(5,6,2,3,4)
    print(x)

数据比较多，就不打印出来呢，反正最后保证降维后升维还能够恢复原来数据顺序。

接下来就是关键代码了：

    xs = torch.randn(8,3, 16, 32, 32)
    conv3d = nn.Conv3d(3, 32, bias=False, kernel_size=3,padding=1)
    conv3d_result = conv3d(xs)
    print(conv3d_result.shape)

    conv2d = nn.Conv2d(3, 32, bias=False, kernel_size=3,padding=1)
    conv3x3_result = []
    for x in zip(xs.unbind(dim=2)):
        x = conv2d(x[0])
        
        conv3x3_result.append(x)
    conv3x3_result = torch.stack(conv3x3_result, dim=2)

    b,c,d,h,w = conv3x3_result.shape
    conv1d = nn.Conv1d(32, 32, bias=False, kernel_size=3,padding=1)

    x = conv3x3_result.transpose(2,4)
    x = x.transpose(2,3)
    x = x.reshape(b,c,d*h*w)
    x = conv1d(x)
    x = x.reshape(b,c,w,h,d)
    x = x.transpose(3,4)
    x = x.transpose(2,3)
    x = x.reshape(b,c,d,h,w)

这样就把5维数据3D卷积来处理转换成了2D和1D卷积处理了，目前来看理论上应该没问题，具体是否正确还需要验证。