Ncnn: quantized int8 storage and operation

Created on 4 Jan 2018 · 5Comments · Source: Tencent/ncnn

enhancement

Source

nihui

Most helpful comment

[x] int8/uint8 storage support in Mat
[x] Quantize Dequantize layer
[x] Integer convolution with bias
[x] ReLU6 layer (implemented as Clip)
[ ] tensorflow quantized mobilenet model
[ ] arm neon optimization

nihui on 5 Jan 2018

👍4

All 5 comments

[x] int8/uint8 storage support in Mat
[x] Quantize Dequantize layer
[x] Integer convolution with bias
[x] ReLU6 layer (implemented as Clip)
[ ] tensorflow quantized mobilenet model
[ ] arm neon optimization

nihui on 5 Jan 2018

👍4

This project seems like a good reference to quantize a NN model under embedded environment:
https://github.com/ARM-software/ML-KWS-for-MCU/blob/master/Deployment/Quant_guide.md

yanghanxy on 11 Jun 2018

Hi, I wonder if the computing of each layer is fixed point data type. As far as I know, the ristretto implements their dynamic fixed point data type by simulation. They just tranfer the data to fixed point and assgin back using float. Thus the real computing is still based on float.