debug and cleanup (cherrypicked from #238)

CMakeLists.txt

@@ -8,6 +8,7 @@ INCLUDE(cmake/3rdparty.cmake)
 
 OPTION(WITH_TRACE "Build with tracing library." OFF)
 OPTION(BUILD_FAKE "Build fake target without CUDA libraries." OFF)
+OPTION(ENABLE_CUDA "Build with CUDA support." ON)
 
 IF(WITH_TRACE)
     MESSAGE(STATUS "[OPENPOSE-PLUS] Trace Enabled.")

cmake/3rdparty.cmake

@@ -3,28 +3,34 @@ INCLUDE(ExternalProject)
 SET(THIRDPARTY_PREFIX ${CMAKE_SOURCE_DIR}/3rdparty)
 
 SET(STDTRACER_GIT_URL
-        https://github.com/lgarithm/stdtracer.git
-        CACHE STRING "URL for clone stdtracer")
+    https://github.com/lgarithm/stdtracer.git
+    CACHE STRING "URL for clone stdtracer")
 
 EXTERNALPROJECT_ADD(
-        stdtracer-repo
-        GIT_REPOSITORY ${STDTRACER_GIT_URL}
-        GIT_TAG c2c7e99fbbd6ead266611c2497cc1ed88c63b46c
-        PREFIX ${THIRDPARTY_PREFIX}
-        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${THIRDPARTY_PREFIX} -DBUILD_TESTS=0
-        -DBUILD_EXAMPLES=0 -DCMAKE_CXX_FLAGS=-fPIC)
+    stdtracer-repo
+    LOG_DOWNLOAD ON
+    LOG_INSTALL ON
+    LOG_CONFIGURE ON
+    GIT_REPOSITORY ${STDTRACER_GIT_URL}
+    GIT_TAG c2c7e99fbbd6ead266611c2497cc1ed88c63b46c
+    PREFIX ${THIRDPARTY_PREFIX}
+    CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${THIRDPARTY_PREFIX} -DBUILD_TESTS=0
+               -DBUILD_EXAMPLES=0 -DCMAKE_CXX_FLAGS=-fPIC)
 
 SET(STDTENSOR_GIT_URL
-        https://github.com/lgarithm/stdtensor.git
-        CACHE STRING "URL for clone stdtensor")
+    https://github.com/lgarithm/stdtensor.git
+    CACHE STRING "URL for clone stdtensor")
 
 EXTERNALPROJECT_ADD(
-        stdtensor-repo
-        GIT_REPOSITORY ${STDTENSOR_GIT_URL}
-        GIT_TAG v0.6.1
-        PREFIX ${THIRDPARTY_PREFIX}
-        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${THIRDPARTY_PREFIX} -DBUILD_TESTS=0
-        -DBUILD_EXAMPLES=0)
+    stdtensor-repo
+    LOG_DOWNLOAD ON
+    LOG_INSTALL ON
+    LOG_CONFIGURE ON
+    GIT_REPOSITORY ${STDTENSOR_GIT_URL}
+    GIT_TAG v0.6.1
+    PREFIX ${THIRDPARTY_PREFIX}
+    CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${THIRDPARTY_PREFIX} -DBUILD_TESTS=0
+               -DBUILD_EXAMPLES=0)
 
 INCLUDE_DIRECTORIES(${THIRDPARTY_PREFIX}/include)
 
@@ -37,4 +43,4 @@ LINK_DIRECTORIES(${THIRDPARTY_PREFIX}/lib)
 FUNCTION(ADD_GLOBAL_DEPS target)
     ADD_DEPENDENCIES(${target} all-external-projects)
     TARGET_LINK_LIBRARIES(${target} stdtracer)
-ENDFUNCTION()
+ENDFUNCTION()

cmake/cuda.cmake

@@ -0,0 +1,15 @@
+IF(IS_DIRECTORY ${CUDA_TOOLKIT_ROOT_DIR})
+    MESSAGE("Using CUDA_TOOLKIT_ROOT_DIR=${CUDA_TOOLKIT_ROOT_DIR}")
+    INCLUDE_DIRECTORIES(${CUDA_TOOLKIT_ROOT_DIR}/include)
+    LINK_DIRECTORIES(${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+ELSE()
+    FIND_PACKAGE(CUDA REQUIRED)
+    INCLUDE_DIRECTORIES(${CUDA_INCLUDE_DIRS})
+    EXECUTE_PROCESS(
+        COMMAND arch
+        COMMAND tr -d '\n'
+        OUTPUT_VARIABLE ARCH)
+    SET(CUDA_RT /usr/local/cuda/targets/${ARCH}-linux)
+
+    LINK_DIRECTORIES(${CUDA_RT}/lib)
+ENDIF()

cmake/openpose-plus.cmake

@@ -1,17 +1,8 @@
+INCLUDE(cmake/cuda.cmake)
 FIND_PACKAGE(OpenCV)
 
 ADD_DEFINITIONS(-Ofast -march=native)
 
-FIND_PACKAGE(CUDA REQUIRED)
-INCLUDE_DIRECTORIES(${CUDA_INCLUDE_DIRS})
-EXECUTE_PROCESS(
-    COMMAND arch
-    COMMAND tr -d '\n'
-    OUTPUT_VARIABLE ARCH)
-SET(CUDA_RT /usr/local/cuda/targets/${ARCH}-linux)
-
-LINK_DIRECTORIES(${CUDA_RT}/lib)
-
 ADD_LIBRARY(openpose-plus src/cudnn.cpp src/paf.cpp src/uff_runner.cpp)
 TARGET_LINK_LIBRARIES(
     openpose-plus

configure

@@ -1,12 +1,9 @@
 #!/bin/bash
 set -e
 
-CMAKE_FLAGS=
-
-add_cmake_flag() {
-    echo "usinig $1=$2"
-    CMAKE_FLAGS="$CMAKE_FLAGS -D$1=$2"
-}
+if [ -d /usr/local/cuda ]; then
+    CUDA_HOME=/usr/local/cuda
+fi
 
 parse_args() {
     for i in "$@"; do
@@ -20,26 +17,40 @@ parse_args() {
         --build-fake)
             BUILD_FAKE=1
             ;;
+        --with-cuda=*)
+            CUDA_HOME="${i#*=}"
+            echo "configure --with-cuda=$CUDA_HOME"
+            ;;
         *)
-            echo "unknown argument $1"
+            echo "unknown argument $i"
             exit
             ;;
         esac
     done
 }
 
+CMAKE_FLAGS=
+
+add_cmake_flag() {
+    echo "using $1=$2"
+    CMAKE_FLAGS="$CMAKE_FLAGS -D$1=$2"
+}
+
 add_cmake_flags() {
     if [ "${BUILD_FAKE}" == 1 ]; then
         add_cmake_flag BUILD_FAKE ON
     fi
     if [ "${WITH_TRACE}" == 1 ]; then
         add_cmake_flag WITH_TRACE ON
     fi
-
-    add_cmake_flag CMAKE_EXPORT_COMPILE_COMMANDS ON
+    if [ ! -z "${CUDA_HOME}" ]; then
+        add_cmake_flag CUDA_TOOLKIT_ROOT_DIR $CUDA_HOME
+        add_cmake_flag CMAKE_CUDA_COMPILER $CUDA_HOME/bin/nvcc
+    fi
     add_cmake_flag CMAKE_BUILD_TYPE Release
-    add_cmake_flag EXECUTABLE_OUTPUT_PATH $(pwd)/bin
-    add_cmake_flag LIBRARY_OUTPUT_PATH $(pwd)/lib
+    add_cmake_flag CMAKE_EXPORT_COMPILE_COMMANDS ON
+    add_cmake_flag EXECUTABLE_OUTPUT_PATH $PWD/bin
+    add_cmake_flag LIBRARY_OUTPUT_PATH $PWD/lib
 }
 
 parse_args $@

examples/example_batch_detector.cpp

@@ -22,25 +22,36 @@ DEFINE_bool(use_f16, false, "Use float16.");
 DEFINE_bool(flip_rgb, true, "Flip RGB.");
 
 // input flags
+DEFINE_string(data_dir, "../../data/media", "path to data folder");
 DEFINE_string(image_files,
-              "../../data/media/COCO_val2014_000000000192.jpg,"
-              "../../data/media/COCO_val2014_000000000459.jpg,"
-              "../../data/media/COCO_val2014_000000000415.jpg,"
-              "../../data/media/COCO_val2014_000000000564.jpg,"
-              "../../data/media/COCO_val2014_000000000294.jpg,"
-              "../../data/media/COCO_val2014_000000000623.jpg,"
-              "../../data/media/COCO_val2014_000000000357.jpg,"
-              "../../data/media/COCO_val2014_000000000488.jpg,"
-              "../../data/media/COCO_val2014_000000000589.jpg,"
-              "../../data/media/COCO_val2014_000000000474.jpg,"
-              "../../data/media/COCO_val2014_000000000338.jpg,"
-              "../../data/media/COCO_val2014_000000000569.jpg,"
-              "../../data/media/COCO_val2014_000000000544.jpg,"
-              "../../data/media/COCO_val2014_000000000428.jpg,"
-              "../../data/media/COCO_val2014_000000000536.jpg,"
-              "../../data/media/COCO_val2014_000000000395.jpg",
+              "COCO_val2014_000000000192.jpg,"
+              "COCO_val2014_000000000459.jpg,"
+              "COCO_val2014_000000000415.jpg,"
+              "COCO_val2014_000000000564.jpg,"
+              "COCO_val2014_000000000294.jpg,"
+              "COCO_val2014_000000000623.jpg,"
+              "COCO_val2014_000000000357.jpg,"
+              "COCO_val2014_000000000488.jpg,"
+              "COCO_val2014_000000000589.jpg,"
+              "COCO_val2014_000000000474.jpg,"
+              "COCO_val2014_000000000338.jpg,"
+              "COCO_val2014_000000000569.jpg,"
+              "COCO_val2014_000000000544.jpg,"
+              "COCO_val2014_000000000428.jpg,"
+              "COCO_val2014_000000000536.jpg,"
+              "COCO_val2014_000000000395.jpg",
               "Comma separated list of paths to image.");
 
+std::vector<std::string> get_input_files(const std::string &dir,
+                                         const std::string &names)
+{
+    std::vector<std::string> paths;
+    for (const auto &name : split(names, ',')) {
+        paths.push_back(dir + "/" + name);
+    }
+    return paths;
+}
+
 int main(int argc, char *argv[])
 {
     TRACE_SCOPE(__func__);
@@ -49,7 +60,8 @@ int main(int argc, char *argv[])
     // TODO: derive from model
     const int f_height = FLAGS_input_height / 8;
     const int f_width = FLAGS_input_width / 8;
-    auto files = repeat(split(FLAGS_image_files, ','), FLAGS_repeat);
+    const auto files = repeat(
+        get_input_files(FLAGS_data_dir, FLAGS_image_files), FLAGS_repeat);
 
     std::unique_ptr<pose_detector> pd(create_pose_detector(
         FLAGS_model_file, FLAGS_input_height, FLAGS_input_width, f_height,

scripts/download-pretrained-inf-models.sh

@@ -10,6 +10,7 @@ cd ../data/models
 
 echo "Installing Pretrained VGG uff model file..."
 if [ ! -f hao28-600000-256x384.uff ]; then
-    curl -LOJ https://github.com/tensorlayer/pretrained-models/trunk/models/openpose-plus/hao28-600000-256x384.uff
+    URL=https://media.githubusercontent.com/media/tensorlayer/pretrained-models/master/models/openpose-plus/hao28-600000-256x384.uff
+    curl -vLOJ $URL
 fi
 echo "Installation completed!"

src/logger.h

@@ -1,5 +1,4 @@
 #pragma once
-
 #include <fstream>
 #include <iostream>
 #include <string>
@@ -9,6 +8,14 @@
 // Logger for TensorRT info/warning/errors
 class Logger : public nvinfer1::ILogger
 {
+    std::string log_file = "tensorrt.log";
+    std::fstream fs;
+
+    void log_with(std::ostream &os, const char *prefix, const char *msg) const
+    {
+        os << prefix << msg << std::endl;
+    }
+
   public:
     Logger() : Logger(Severity::kINFO) {}
 
@@ -21,31 +28,28 @@ class Logger : public nvinfer1::ILogger
     {
         // suppress messages with severity enum value greater than the
         // reportable
-        if (severity > reportableSeverity) return;
+        if (severity > reportableSeverity) { return; }
 
         switch (severity) {
         case Severity::kINTERNAL_ERROR:
-            fs << "INTERNAL_ERROR: ";
+            log_with(fs, "INTERNAL_ERROR: ", msg);
+            log_with(std::cerr, "INTERNAL_ERROR: ", msg);
             break;
         case Severity::kERROR:
-            fs << "ERROR: ";
+            log_with(fs, "ERROR: ", msg);
+            log_with(std::cerr, "ERROR: ", msg);
             break;
         case Severity::kWARNING:
-            fs << "WARNING: ";
+            log_with(fs, "WARNING: ", msg);
             break;
         case Severity::kINFO:
-            fs << "INFO: ";
+            log_with(fs, "INFO: ", msg);
             break;
         default:
-            fs << "UNKNOWN: ";
+            log_with(fs, "UNKNOWN: ", msg);
             break;
         }
-        fs << msg << std::endl;
     }
 
     Severity reportableSeverity{Severity::kWARNING};
-
-  private:
-    std::string log_file = "tensorrt.log";
-    std::fstream fs;
 };

src/uff_runner.cpp

@@ -172,8 +172,7 @@ void uff_runner_impl::createBuffers_(int batch_size)
         std::cout << "binding " << i << ":"
                   << " name: " << name << " type" << to_string(dtype)
                   << to_string(dims) << std::endl;
-        buffers_.emplace_back(
-            cuda_buffer_t(batch_size, volume(dims) * elementSize(dtype)));
+        buffers_.emplace_back(batch_size, volume(dims) * elementSize(dtype));
     }
 }