模型加速

2019-05-26 11:48:02 浏览数 (3)

CLIP-Q: Deep Network Compression Learning by In-Parallel Pruning-Quantization CVPR2018 http://www.sfu.ca/~ftung/ 裁剪和量化一体化框架

本文的思路比较简单,裁剪 量化一体训练模型分三个步骤: 1) Clipping 裁剪,将网络中的权重系数值接近0 的权重全部置零,当然这种置零是临时性的,后面的训练迭代根据实际情况调整。 这里的阈值自适应确定,(model the objective function as a Gaussian process) 2)Partitioning 切分, partition the non-clipped portion of the 1-D axis of weight values into quantization intervals,这里我们使用了 linear (uniform) partitioning ,也可以使用其他自适应切分 如 weighted entropy 3)Quantizing 量化 update the quantization levels the discrete values that the weights are permitted to take in the compressed network

Experiments

11

0 人点赞