MXNet2Caffe2trt: Convert MXNet model to Caffe model,Then Convert Caffe model to TRT model

• first:Convert Mxnet model to Caffe model
• then :Convert Caffe model to Tensorrt model

Mxnet2Caffe

• json2prototxt.py prototxt_basic.py Read mxnet_json file and converte to prototxt.
• mxnet2caffe.py Read mxnet_model params_dict and converte to .caffemodel.
• mxnet_caffe_model_test.py Compare the outputs of Caffe model and TRT model.
• mxnet_t2t.py Convert training model to inference model.Note you should change the file for your model.
• Usage
  • First : Using json2prototxt.py to convert json file to prototxt. Using json2prototxt.py -h to get the args
  • Then : Using mxnet2caffe.py to convert params file to caffemodel. Using mxnet2caffe.py -h to get the args
  • Final : Using mxnet_caffe_model_test.py to compare the model outputs. Using mxnet_caffe_model_test.py -h to get the args
  • Note:Uisng Netron(https://github.com/lutzroeder/Netron) to see the model structure

caffe2trt_int8

• caffe2trt_int8.h Convert Caffe model to TRT Int8 model.
• calibrator.h Get the Calibratir for int8 model.
• CopyPlugin.h Example for Add Plugin.
• Note
  • From https://blog.csdn.net/qq_32043199/article/details/81119357, you can get the detail about float32 to int8
  • In toint8.cpp, you can see how to use these code to convert the Caffe model to TRT Int8 model

caffe2trt

• caffe2trt_half.h Convert Caffe model to TRT model. Note half mode will be used if device supports(NVIDIA Turing GPU architecture)

trt_inference

• trt2engine.h Load TRT model and do model Inference. Note just one input and one output are supported now.

Build

• docker pull sunny820828449/mxnet2trt
• docker run -t -i sunny820828449/mxnet2trt
• git clone http://git.sysop.bigo.sg/feature-db/mxnet_caffe_trt.git
• Enter mxnet_caffe_trt and bash mk.sh