Implement R-FCN in TensorRT, How to write the psroi_pooling layer?

Hi, every one, I am trying to implement R-FCN on TensorRT which is more faster and more accurate then faster-rcnn, but since the sample-faster-rcnn are hide the implementation of ROI-fused-pooling layer so that we have no way to hack that into R-FCN.

I checked the documentation, I am able to write a plugin to do that, but question is how to get the middle output of network?

in createFasterRCNNPlugin(), this method even not send a middle data into it. what is the logic exactly??


Please check our face_recognition sample for more information.
Here is an enqueue function to demonstrate the access of middle data.


Thank you Pal. But I see the example, I implement all the methods of IPlugin and I can do this step:

std::cout << "Begin parsing model..." << std::endl;
    const IBlobNameToTensor* blobNameToTensor = parser->parse(deployFile.c_str(),
    std::cout << "End parsing model..." << std::endl;

How ever afer this when call ICudaEngine* engine = builder->buildCudaEngine(*network); it’s core dumped. Seems it’s the size or dim issue. But I logged all the layer init function, I think at least it should has some log, but nothing. I can not debug this. Any help to debug it???


Try to output more log information here:


Hi,I want to know what networks the face-recognition sample used.And,to implement R-FCN in TensorRT,I want to know how to write the psroi_pooling layer and write a plugin?