Foreword

NVIDIA in open source projects,onnx-tensorrtVersion is v5.0, currently of 18.04 TX2 system Cuda10, Tensorrt5.0.26 currently only supports this version.
tensorrt we can see how the model into a sequence of ONNX model tensorrt support. Source code, we can see there are four Op: Split, FancyActivation, InstanceNormalization, ResizeNearest are later registered support, not tensorrt internal supports.
so we register their own custom Op, op may need to consult on how to achieve these registration functions.

Register your own custom Op: dcn_v2_forward

DCN V2 is deformable convolution of the latest version implements, DCN has many excellent characteristics, in particular target detection algorithm, adding DCN can effectively improve the detection rate. DCNV2 the current version of the source https://github.com/CharlesShang/DCNv2, currently has Cuda accelerate, so we DCN V2 as a custom Op registered to ONNX-Tensorrt in.

Sign up concrete steps

1. Write source code dcn_v2_forward.hpp and dcn_v2_forward.cu, there is dcn_v2_forward of Op concrete realization, heavy onnx2trt :: Plugin, note that this is different from the tensorrt caffe registered layer or layers is tensorrt register themselves.

class dcn_v2_forwardPlugin final : public onnx2trt::Plugin 

Specific reference may be written in step Split, FancyActivation, InstanceNormalization, to achieve the ResizeNearest.
3. In builtin_plugins.cpp rearmost, as registration of the following is added as a dcn_v2_forward.

REGISTER_BUILTIN_PLUGIN("FancyActivation",       FancyActivationPlugin);
REGISTER_BUILTIN_PLUGIN("ResizeNearest",         ResizeNearestPlugin);
REGISTER_BUILTIN_PLUGIN("Split"        ,         SplitPlugin);
REGISTER_BUILTIN_PLUGIN("InstanceNormalization", InstanceNormalizationPlugin);
REGISTER_BUILTIN_PLUGIN("dcn_v2_forward",        dcn_v2_forwardPlugin);

REGISTER_BUILTIN_NVPLUGIN("Concat", ConcatPlugin);

3. The core of the steps, builtin_op_importers.cpp used DEFINE_BUILTIN_OP_IMPORTER added using their own registration Op, only take effect here, use this procedure onnx2trt time to find our own custom Op.