卷积核为1×1的conv转为矩阵乘运算，pytorch实现

主要就是ci和ci相乘，然后co方向加上bias。

代码如下：

import torch

def test_conv_to_matmul(n, h, w, ci, co):
    conv = torch.nn.Conv2d(ci, co, (1, 1)).cuda()
    input = torch.randn(n, ci, h, w).cuda()
    output = conv(input)
    # using matmul
    output_matmul = torch.matmul(conv.weight.reshape((co, ci)), \
            input.permute(0, 2, 3, 1).reshape((-1, ci)).permute((1,0))).permute(1, 0) + conv.bias
    print(torch.sum(output - output_matmul.reshape(n, h, w, co).permute(0, 3, 1, 2)), flush = True)

if __name__=="__main__":
    test_conv_to_matmul(10, 20, 30, 6, 8)