INCREMENTAL NETWORK QUANTIZATION: TOWARDS LOSSLESS CNNS WITH LOW-PRECISION WEIGHTS

目標:

將pre-trained full-precision (32-bit floating-point) CNN model 轉成低精度(2的次方或1)

好處: 乘法可以轉換成 binary bit shift operations (FPGA)。

大部分的量化權重法都是針對所有權重的,但是沒有考慮不同權重的重要性不同。

本篇作法: weight partition, group-wise quantization and re-training

把CNN model 每一層的權重分成兩類,一類權重被量化成低精度 (0或者2的倍數),一類權重保持原本 精度,完成後重新訓練,更新參數。接著,再作一次上述步驟,直到所有權重都被量化。

http://hongbomin.com/2017/02/17/paper-reading-INQ/

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.