To accelerate mobileNet-ssd with tensorRT
TensorRT-Mobilenet-SSD can run 50fps on jetson tx2
Requierments:
1.tensorRT4
2.cudnn7
3.opencv
Run:
cmake .
./build/bin/mobileNet
Reference:
https://github.com/saikumarGadde/tensorrt-ssd-easy
https://github.com/chuanqi305/MobileNet-SSD
I replaced depthwise with group_conv,because group_conv has been optimized in cudnn7
I retrianed mobileNet-SSD,my number of classfication is 5
TODO:
1.serializing model will cost a lot of time, if could save the result of serializing model to disk, it will be more convenient
2.I think the cost of inference is not good,I want to change the architecture.If I succeeded,I will push it
3.I found that same code get different result in gtx1080 and tx2. WHY?????
4.There is a problem，the code runs on different platforms, get different result.
All the problems mentioned above have been solved
There are bugs in many related code on github
If you want to implement tensorRT-mobileNet, it is enough to refer to me.