Skip to content

tenglike1997/onnxruntime-projects

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

31 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

onnxruntime projects

Introduction

This repository include codes for some onnxruntime projects,such as classification, segmentation, detection, style transfer and super resolution.

Onnxruntime

ONNX Runtime is a performance-focused complete scoring engine for Open Neural Network Exchange (ONNX) models, with an open extensible architecture to continually address the latest developments in AI and Deep Learning. In my repository,onnxruntime.dll have been compiled. You can download it and see specific information about onnxruntime in https://github.com/microsoft/onnxruntime.

Projects

The programming language is C++ and The platform is Visual Studio. I have finished some projects based on onnxruntime official samples. The link have been mentioned afore. Also, you can download some onnx models in https://github.com/onnx/models. If necessary,you can see the structure onnx models in https://lutzroeder.github.io/netron/.

Windows
Network Classes Input resolution Batch size Iterations CPU Running time GPU Running time TRT Running time*
MobileNet 1000 224x224 1 1000 19.56s 4.15s 1.05s
ERFNet 4 640x480 1 1000 >100s 12.93s 5.6s
Tiny_YOLOv2 20 416x416 1 1000 40.64s 2.97s 1.92s
Super Resolution with sub-pixel CNN - 224x224 1 1000 34.14s 1.79s 1.14s
Fast Neural Style Transfer - 224x224 1 1000 87.99s 4.64s -
Ubuntu
Network Classes Input resolution Batch size Iterations CPU Running time GPU Running time TRT Running time*
MobileNet 1000 224x224 1 1000 20.09s 4.24s 0.79s
ERFNet 4 640x480 1 1000 >100s 13.56s 4.90s

*The TensorRT engine is compiled with FP16 settings. Just add "trt_builder->setFp16Mode(true);" to 339 line of tensorrt_execution_provider.cc, if you build libonnxruntime yourself.

**This experiment is implemented on NVIDIA 2080Ti.

Classification


The onnx model is moblienet. You can download it in the link mentioned afore.

Segmentation


The onnx model is our trained erfnet. We use specific datasets to train erfnet.

Detection


The onnx model is Tiny YOLOv2.You can download it in the link mentioned afore.

Style transfer


The onnx model is Fast Neural Style Transfer. You can download it in the link mentioned afore.

Super resolution


The onnx model is Super Resolution with sub-pixel CNN. You can download it in the link mentioned afore.