search

zhreshold / mxnet-ssd.cpp

C++ object detection module for mxnet-ssd
MIT License
97 stars 47 forks source link

400ms on gpu/1060 which is much slow than yolo #10

Open anguoyang opened 7 years ago

anguoyang commented 7 years ago

i have tested on your demo on gpu mode, it took about 400ms for 1 image, which is much slow than yolo(<100ms).

nopattern commented 7 years ago
  1. The default config is not using cuda and cudnn. you should check your config.
  2. As in my case,the first time of running will take more time than usual. you can test more.
anguoyang commented 7 years ago

hi,thank you, I have checked the pre-defined values before the main, and debug to see the device type and device id, they are OK. and i also added clock related lines before and after the detect function to calculate the duration only on the detection. is there any other configurations?

nopattern commented 7 years ago

you can just use normal mxnet-ssd if you want to run on gtx1060, Why to use this version? This version seems that it supports cuda but doesn't support cudnn .

anguoyang commented 7 years ago

because I want to use mxnet in c++

zhreshold commented 7 years ago

How did you get GPU to work using this repo? I think the makefile don't support GPU yet.

anguoyang commented 7 years ago

I modify on visual studio to run it, GPU is working,however, cudnn is not working

zhreshold commented 7 years ago

Can you profile the runtime?

anguoyang commented 7 years ago

profile?

nopattern commented 7 years ago

The default config about cuda is #define MSHADOW_USE_CUDA 0, As @threshold ,the default seems don't support cuda. You can use c_predict_api.h to use gtx1060 in c++ .

pelre commented 7 years ago

@anguoyang how did you modify it on visual studio to work with GPU? I'd gladly accept any pointers to be able to do the same

aa12356jm commented 6 years ago

@zhreshold i can not get gpu to work ,could you give me some tips to get gpu to work? thanks