Plugins for Pytorch & TensorRT
It's a personal backup repository, but I will try my best to help you understand how to use them.
| Now it contains: | Plugins | Support |
|---|---|---|
| DCNv2 | pytorch1.5+/tensorrt7 | |
| yolov5 detection | tensorrt7/tensorrt8 |
| Models | Support |
|---|---|
| mobilenetv3-centernet | pytorch1.5+/tensorrt7 |
| yolov5 (integrated with detection and nms) | tensorrt7/tensorrt8 |
Installation
Pytorch
cd pytorch
python setup.py install --user(optioanl)
cd ../examplespython
from dcn_v2_wrapper import DeformableConv2DLayer as DCN
TensorRT
-
Copy plugin folders from
tensorrttoNVIDIA/TensorRT/plugin -
Add relative head file and initializePlugin() to InferPlugin.cpp at proper place, for example
#include "dcnv2Plugin.h"
#include "yoloPlugin.h"
initializePlugin<nvinfer1::plugin::DCNv2PluginCreator>(logger, libNamespace);initializePlugin<nvinfer1::plugin::YoloPluginCreator>(logger, libNamespace);
-
Add name of plugin folder to
PLUGIN_LISTSin CMakeLists.txt -
Build and use
libnvinfer_plugin.sofollowing offical introduction.
EXAMPLES
There are two pytorch2tensorrt transfer scripts in examples to show how these plugins work.
Evaluation
cd examples
python mbv3_centernet_trt7.py
The evaluation output is as follow which are mean values of hm, wh, reg
