search

AlexeyAB / yolo2_light

Light version of convolutional neural network Yolo v3 & v2 for objects detection with a minimum of dependencies (INT8-inference, BIT1-XNOR-inference)
MIT License
302 stars 116 forks source link

is that possible to use full INT8 during tiny-yolo-v3 #46

Open jasonwu1977 opened 5 years ago

jasonwu1977 commented 5 years ago

Hi @AlexeyAB I would like to know is that possible to use full INT8 during all yolo-light2

  1. after pre-processing layer is that possible to convert it to INT8 before entering the first layer
  2. can i use INT8/INT16 bias
  3. can I save output to 8 bit before next layer (of course befoer yolo layer it still float32)
jasonwu1977 commented 5 years ago

@AlexeyAB Answer for myself All 3 questions are yes, and i have tested with CPU

I want to run quantization with pure CPU mode. And When I change the Makefile to let yolo2-light running on CPU mode. the mAP drops to very low, after I trace the code, i noticed the quantization of CUDNN mode is different with running GPU mode (some layer don't quantize in CUDNN mode)

Now the question is, on yolov2 or yolov3, on CUDNN examples, the size=1 & stride=2 layer doesn't do the quantize, but is it possible to do quantize with CPU mode? And if i want to change the CPU to support all convolution layer quantization, where should i moidfy?

jason4345099 commented 4 years ago

@jasonwu1977 hi Jason, I met similar issue as you, could you can talk it more about it through my qq 494529371? big thanks.

Joejwu commented 2 years ago

Have you sloved it out yet? or anyone else know?