Incremental-Network-Quantization
Incremental-Network-Quantization copied to clipboard
why use 1 bit to store 0
For example, for 4bit why don't we just use 15 to form quantized region, and another one such as 0000 to represent 0. Will unbalanced region influence INQ result?
yes, for 4 bits, you can use 15 no-zero values + 0, unbalanced quantization doesn't influence INQ result, specifically, the INQ is very flexible,you can adjust bit-width according your task, for example , you can throw away 0,then, the 5 bits in INQ paper can be replaced by 4 bits.
- template double weightCluster
(int weight,int M); - template double weightCluster_zero
(float weight,int M); why this two method is all the same? the second one shoud contain 0? @Zhouaojun