1. 论文思想

一维滤过器。将三维卷积分解成三个一维卷积。convolution across channels(lateral), vertical and horizontal direction.

论文笔记—Flattened convolution neural networks for feedforward acceleration ### 2. 计算量对比 论文笔记—Flattened convolution neural networks for feedforward acceleration

变换后计算量:
论文笔记—Flattened convolution neural networks for feedforward acceleration
对比:
论文笔记—Flattened convolution neural networks for feedforward acceleration
论文笔记—Flattened convolution neural networks for feedforward acceleration

3. 总结

因为spatial convolution会带来大量的参数以及是非常耗时的,本文将三维卷积分解成了三个一维的卷积,极大的减少了计算量。其实,本文也引入了不对称卷积,再后来也证实了这种不对称卷积Nx1和1xN,对准确率是有提升的。

相关文章: