First You shall compile the cuda_gpu for c++

cd utils  
bash do.sh  
cp ./build/libgpu_nms.so ../  

Second compile your refinedet API

cd ../
bash do.sh  

This model for 2 detector, if you want to add more.
U only change a little from my code.
You must train your refinedet in your pytorch,and import the torch script.
U must give the anchor priors about U model and save to *.csv

U must put your test_img into JPEGImages.do follow: U give a result dir for your test output and get the filename save into file_name.txt.

mkdir JPEGImages  
mkdir result

./build/launch [file_name.txt][U don't need][anchor.csv][Torch_model]  

IF U have any question Email me.