Current Result Document :
| 한글제목(Korean Title) |
Channel-wise Mixed precision considering quantization sensitivity and layer characteristics |
| 영문제목(English Title) |
Channel-wise Mixed precision considering quantization sensitivity and layer characteristics |
| 저자(Author) |
Yongjoo Lee
Mincheal Kang
Jiwon Seo
|
| 원문수록처(Citation) |
VOL 49 NO. 02 PP. 0580 ~ 0582 (2022. 12) |
한글내용
(Korean Abstract) |
|
영문내용
(English Abstract) |
Quantization is one of the model compression techniques to reduce computation and memory footprint efficiently and improve inference speed by representing a continuous real number as a discrete value. Recently, research has been conducted to find an appropriate bitwidth by analyzing layer-wise or channel-wise characteristics with reinforcement learning based on a mixed precision technique that uses two or more types of bits mixed instead of a single bit. Applying reinforcement learning can minimize quantization errors to maximize performance but requires high costs for exploring numerous bit combinations. In this paper, using two data types, INT4 and INT8, 1) the bit combination is updated while gradually increasing the ratio of INT8 data type through sensitivity analysis on quantization for each channel of the convolution layer, 2) QAT (Quantization Aware Training) ) technique to keep model performance degradation to a minimum. As a result of the experiment, the optimal bit combination was found through a searching process of up to 20 times, and the performance was maintained as in the case of setting all convolution layer data types as INT8. |
| 키워드(Keyword) |
|
| 파일첨부 |
PDF 다운로드
|
