• Stars
    star
    155
  • Rank 240,864 (Top 5 %)
  • Language
    C++
  • License
    MIT License
  • Created over 4 years ago
  • Updated 10 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Useful tensorrt plugin. For pytorch and mmdetection model conversion.

Amirstan_plugin

Amirstan plugin contains some useful TensorRT plugins. These plugins are used to support other projects:

Requirement

  • TensorRT >= 7.0.0.11 (Support TensorRT 8 now)

Installation

From sources

  • clone the repo and create build folder

    git clone --depth=1 https://github.com/grimoire/amirstan_plugin.git
    cd amirstan_plugin
    git submodule update --init --progress --depth=1
    mkdir build
    cd build
  • build the project with:

    • plugins only

      cmake -DTENSORRT_DIR=${path_to_tensorrt} ..
      make -j$(nproc)
    • or with DeepStream support

      cmake -DTENSORRT_DIR=${path_to_tensorrt} -DWITH_DEEPSTREAM=true -DDeepStream_DIR=${path_to_deepstream} ..
      make -j$(nproc)
  • set the envoirment variable(in ~/.bashrc):

export AMIRSTAN_LIBRARY_PATH=<amirstan_plugin_root>/build/lib

Using Conan

  • Install Conan:

    pip install conan
  • Register grimoire's Conan remote:

    conan remote add grimoire https://grimoire.jfrog.io/artifactory/api/conan/grimoire-conan
  • Add a conanfile.txt file to your project's root with the following content:

    [requires]
    amirstan_plugin/0.5.0
    
    [generators]
    cmake
  • Additionaly, you can add a few options under the [options] section to configure your build:

    • tensorrt_dir: path where TensorRT is located. Default ~/SDK/TensorRT.
    • with_deepstream: whether to compile with deepstream support. Default False.
    • deepstream_dir: path where deepstream is located. Default /opt/nvidia/deepstream/deepstream
    • cub_root_dir: Default ./third_party/cub
    • cuda_arch: list of CUDA architectures to compile for. Default 61;62;70;72;75;80;86

    For example, to use a custom TensorRT dir and compile for a specific CUDA architecture:

    [requires]
    amirstan_plugin/0.5.0
    
    [generators]
    cmake
    
    [options]
    amirstan_plugin:tensorrt_dir=/usr/include/x86_64-linux-gnu
    amirstan_plugin:cuda_arch=75
  • Add the following lines to your project root's CMakeLists.txt:

    INCLUDE(${CMAKE_BINARY_DIR}/conanbuildinfo.cmake)
    CONAN_BASIC_SETUP()
  • Add conan libs to the linking stage:

    target_link_libraries(trt_sample PUBLIC ${CONAN_LIBS} ${CUDA_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT} ${TensorRT_LIBRARIES})
  • Compile your project:

    mkdir build
    cd build
    conan install .. -s compiler.libcxx=libstdc++11 --build=missing 
    cmake .. 
    make -j$(nproc)