From f4c384d1b1b4cb854ecc1d311c6f1476db8e42d1 Mon Sep 17 00:00:00 2001 From: makaveli <39617050+makaveli10@users.noreply.github.com> Date: Thu, 6 May 2021 08:35:20 +0530 Subject: [PATCH] add: Scaled yolov4 (#524) * add: mish, yololayer, layers * update: CMake * fix: compile * add: yolov4-csp net def * update: cuda kernel for scaled_yolov4 * increase nms thresh * update: README * add: gen_wts * update: CMake * fix memory leak * Update README.md * Update README.md * Update README.md * Update README.md --- scaled-yolov4/CMakeLists.txt | 37 +++ scaled-yolov4/README.md | 56 ++++ scaled-yolov4/common.hpp | 196 +++++++++++++ scaled-yolov4/gen_wts.py | 23 ++ scaled-yolov4/logging.h | 507 ++++++++++++++++++++++++++++++++++ scaled-yolov4/mish.cu | 196 +++++++++++++ scaled-yolov4/mish.h | 108 ++++++++ scaled-yolov4/utils.h | 39 +++ scaled-yolov4/yololayer.cu | 274 +++++++++++++++++++ scaled-yolov4/yololayer.h | 154 +++++++++++ scaled-yolov4/yolov4_csp.cpp | 516 +++++++++++++++++++++++++++++++++++ 11 files changed, 2106 insertions(+) create mode 100644 scaled-yolov4/CMakeLists.txt create mode 100644 scaled-yolov4/README.md create mode 100644 scaled-yolov4/common.hpp create mode 100644 scaled-yolov4/gen_wts.py create mode 100644 scaled-yolov4/logging.h create mode 100644 scaled-yolov4/mish.cu create mode 100644 scaled-yolov4/mish.h create mode 100644 scaled-yolov4/utils.h create mode 100644 scaled-yolov4/yololayer.cu create mode 100644 scaled-yolov4/yololayer.h create mode 100644 scaled-yolov4/yolov4_csp.cpp diff --git a/scaled-yolov4/CMakeLists.txt b/scaled-yolov4/CMakeLists.txt new file mode 100644 index 0000000..110560d --- /dev/null +++ b/scaled-yolov4/CMakeLists.txt @@ -0,0 +1,37 @@ +cmake_minimum_required(VERSION 2.6) + +project(yolov4) + +add_definitions(-std=c++11) + +option(CUDA_USE_STATIC_CUDA_RUNTIME OFF) +set(CMAKE_CXX_STANDARD 11) +set(CMAKE_BUILD_TYPE Debug) + +find_package(CUDA REQUIRED) + +include_directories(${PROJECT_SOURCE_DIR}/include) +# include and link dirs of cuda and tensorrt, you need adapt them if yours are different +# cuda +include_directories(/usr/local/cuda/include) +link_directories(/usr/local/cuda/lib64) +# tensorrt +include_directories(/usr/include/x86_64-linux-gnu/) +link_directories(/usr/lib/x86_64-linux-gnu/) + +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Ofast -Wfatal-errors -D_MWAITXINTRIN_H_INCLUDED") + +cuda_add_library(myplugins SHARED ${PROJECT_SOURCE_DIR}/yololayer.cu ${PROJECT_SOURCE_DIR}/mish.cu) +target_link_libraries(myplugins nvinfer cudart) + +find_package(OpenCV) +include_directories(${OpenCV_INCLUDE_DIRS}) + +add_executable(yolov4csp ${PROJECT_SOURCE_DIR}/yolov4_csp.cpp) +target_link_libraries(yolov4csp nvinfer) +target_link_libraries(yolov4csp cudart) +target_link_libraries(yolov4csp myplugins) +target_link_libraries(yolov4csp ${OpenCV_LIBS}) + +add_definitions(-O2 -pthread) + diff --git a/scaled-yolov4/README.md b/scaled-yolov4/README.md new file mode 100644 index 0000000..dc94e07 --- /dev/null +++ b/scaled-yolov4/README.md @@ -0,0 +1,56 @@ +# scaled-yolov4 + +The Pytorch implementation is from [WongKinYiu/ScaledYOLOv4 yolov4-csp branch](https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp). It can load yolov4-csp.cfg and yolov4-csp.weights(from AlexeyAB/darknet). + +Note: There is a slight difference in yolov4-csp.cfg for darknet and pytorch. Use the one given in the above repo. + +## Config + +- Input shape `INPUT_H`, `INPUT_W` defined in yololayer.h +- Number of classes `CLASS_NUM` defined in yololayer.h +- FP16/FP32 can be selected by the macro `USE_FP16` in yolov4_csp.cpp +- GPU id can be selected by the macro `DEVICE` in yolov4_csp.cpp +- NMS thresh `NMS_THRESH` in yolov4_csp.cpp +- bbox confidence threshold `BBOX_CONF_THRESH` in yolov4_csp.cpp +- `BATCH_SIZE` in yolov4_csp.cpp + +## How to run + +1. generate yolov4_csp.wts from pytorch implementation with yolov4-csp.cfg and yolov4-csp.weights. + +``` +git clone https://github.com/wang-xinyu/tensorrtx.git +git clone -b yolov4-csp https://github.com/WongKinYiu/ScaledYOLOv4.git +// download yolov4-csp.weights from https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp#yolov4-csp +cp {tensorrtx}/scaled-yolov4/gen_wts.py {ScaledYOLOv4/} +cd {ScaledYOLOv4/} +python gen_wts.py yolov4-csp.weights +// a file 'yolov4_csp.wts' will be generated. +``` + +2. put yolov4_csp.wts into {tensorrtx}/scaled-yolov4, build and run + +``` +mv yolov4_csp.wts {tensorrtx}/scaled-yolov4/ +cd {tensorrtx}/scaled-yolov4 +mkdir build +cd build +cmake .. +make +sudo ./yolov4csp -s // serialize model to plan file i.e. 'yolov4csp.engine' +sudo ./yolov4csp -d ../../yolov3-spp/samples // deserialize plan file and run inference, the images in samples will be processed. +``` + +3. check the images generated, as follows. _zidane.jpg and _bus.jpg +

+ +

+ +

+ +

+ + +## More Information + +See the readme in [home page.](https://github.com/wang-xinyu/tensorrtx) diff --git a/scaled-yolov4/common.hpp b/scaled-yolov4/common.hpp new file mode 100644 index 0000000..5e41a93 --- /dev/null +++ b/scaled-yolov4/common.hpp @@ -0,0 +1,196 @@ +#include +#include +#include +#include +#include + +#include "NvInfer.h" +#include "yololayer.h" +#include "mish.h" + + +using namespace nvinfer1; + +cv::Mat preprocess_img(cv::Mat& img) { + int w, h, x, y; + float r_w = Yolo::INPUT_W / (img.cols*1.0); + float r_h = Yolo::INPUT_H / (img.rows*1.0); + if (r_h > r_w) { + w = Yolo::INPUT_W; + h = r_w * img.rows; + x = 0; + y = (Yolo::INPUT_H - h) / 2; + } else { + w = r_h* img.cols; + h = Yolo::INPUT_H; + x = (Yolo::INPUT_W - w) / 2; + y = 0; + } + cv::Mat re(h, w, CV_8UC3); + cv::resize(img, re, re.size()); + cv::Mat out(Yolo::INPUT_H, Yolo::INPUT_W, CV_8UC3, cv::Scalar(128, 128, 128)); + re.copyTo(out(cv::Rect(x, y, re.cols, re.rows))); + return out; +} + +cv::Rect get_rect(cv::Mat& img, float bbox[4]) { + int l, r, t, b; + float r_w = Yolo::INPUT_W / (img.cols * 1.0); + float r_h = Yolo::INPUT_H / (img.rows * 1.0); + if (r_h > r_w) { + l = bbox[0] - bbox[2]/2.f; + r = bbox[0] + bbox[2]/2.f; + t = bbox[1] - bbox[3]/2.f - (Yolo::INPUT_H - r_w * img.rows) / 2; + b = bbox[1] + bbox[3]/2.f - (Yolo::INPUT_H - r_w * img.rows) / 2; + l = l / r_w; + r = r / r_w; + t = t / r_w; + b = b / r_w; + } else { + l = bbox[0] - bbox[2]/2.f - (Yolo::INPUT_W - r_h * img.cols) / 2; + r = bbox[0] + bbox[2]/2.f - (Yolo::INPUT_W - r_h * img.cols) / 2; + t = bbox[1] - bbox[3]/2.f; + b = bbox[1] + bbox[3]/2.f; + l = l / r_h; + r = r / r_h; + t = t / r_h; + b = b / r_h; + } + return cv::Rect(l, t, r-l, b-t); +} + +float iou(float lbox[4], float rbox[4]) { + float interBox[] = { + std::max(lbox[0] - lbox[2]/2.f , rbox[0] - rbox[2]/2.f), //left + std::min(lbox[0] + lbox[2]/2.f , rbox[0] + rbox[2]/2.f), //right + std::max(lbox[1] - lbox[3]/2.f , rbox[1] - rbox[3]/2.f), //top + std::min(lbox[1] + lbox[3]/2.f , rbox[1] + rbox[3]/2.f), //bottom + }; + + if(interBox[2] > interBox[3] || interBox[0] > interBox[1]) + return 0.0f; + + float interBoxS =(interBox[1]-interBox[0])*(interBox[3]-interBox[2]); + return interBoxS/(lbox[2]*lbox[3] + rbox[2]*rbox[3] -interBoxS); +} + +bool cmp(const Yolo::Detection& a, const Yolo::Detection& b) { + return a.det_confidence > b.det_confidence; +} + +void nms(std::vector& res, float *output, float conf_thresh, float nms_thresh = 0.5) { + int det_size = sizeof(Yolo::Detection) / sizeof(float); + std::map> m; + for (int i = 0; i < output[0] && i < Yolo::MAX_OUTPUT_BBOX_COUNT; i++) { + if (output[1 + det_size * i + 4] <= conf_thresh) continue; + Yolo::Detection det; + memcpy(&det, &output[1 + det_size * i], det_size * sizeof(float)); + if (m.count(det.class_id) == 0) m.emplace(det.class_id, std::vector()); + m[det.class_id].push_back(det); + } + for (auto it = m.begin(); it != m.end(); it++) { + //std::cout << it->second[0].class_id << " --- " << std::endl; + auto& dets = it->second; + std::sort(dets.begin(), dets.end(), cmp); + for (size_t m = 0; m < dets.size(); ++m) { + auto& item = dets[m]; + res.push_back(item); + for (size_t n = m + 1; n < dets.size(); ++n) { + if (iou(item.bbox, dets[n].bbox) > nms_thresh) { + dets.erase(dets.begin()+n); + --n; + } + } + } + } +} + +// TensorRT weight files have a simple space delimited format: +// [type] [size] +std::map loadWeights(const std::string file) { + std::cout << "Loading weights: " << file << std::endl; + std::map weightMap; + + // Open weights file + std::ifstream input(file); + assert(input.is_open() && "Unable to load weight file."); + + // Read number of weight blobs + int32_t count; + input >> count; + assert(count > 0 && "Invalid weight map file."); + + while (count--) + { + Weights wt{DataType::kFLOAT, nullptr, 0}; + uint32_t size; + + // Read name and type of blob + std::string name; + input >> name >> std::dec >> size; + wt.type = DataType::kFLOAT; + + // Load blob + uint32_t* val = reinterpret_cast(malloc(sizeof(val) * size)); + for (uint32_t x = 0, y = size; x < y; ++x) + { + input >> std::hex >> val[x]; + } + wt.values = val; + + wt.count = size; + weightMap[name] = wt; + } + + return weightMap; +} + +IScaleLayer* addBatchNorm2d(INetworkDefinition *network, std::map& weightMap, ITensor& input, std::string lname, float eps) { + float *gamma = (float*)weightMap[lname + ".weight"].values; + float *beta = (float*)weightMap[lname + ".bias"].values; + float *mean = (float*)weightMap[lname + ".running_mean"].values; + float *var = (float*)weightMap[lname + ".running_var"].values; + int len = weightMap[lname + ".running_var"].count; + + float *scval = reinterpret_cast(malloc(sizeof(float) * len)); + for (int i = 0; i < len; i++) { + scval[i] = gamma[i] / sqrt(var[i] + eps); + } + Weights scale{DataType::kFLOAT, scval, len}; + + float *shval = reinterpret_cast(malloc(sizeof(float) * len)); + for (int i = 0; i < len; i++) { + shval[i] = beta[i] - mean[i] * gamma[i] / sqrt(var[i] + eps); + } + Weights shift{DataType::kFLOAT, shval, len}; + + float *pval = reinterpret_cast(malloc(sizeof(float) * len)); + for (int i = 0; i < len; i++) { + pval[i] = 1.0; + } + Weights power{DataType::kFLOAT, pval, len}; + + weightMap[lname + ".scale"] = scale; + weightMap[lname + ".shift"] = shift; + weightMap[lname + ".power"] = power; + IScaleLayer* scale_1 = network->addScale(input, ScaleMode::kCHANNEL, shift, scale, power); + assert(scale_1); + return scale_1; +} + +ILayer* convBnMish(INetworkDefinition *network, std::map& weightMap, ITensor& input, int outch, int ksize, int s, int p, int linx) { + Weights emptywts{DataType::kFLOAT, nullptr, 0}; + IConvolutionLayer* conv1 = network->addConvolutionNd(input, outch, DimsHW{ksize, ksize}, weightMap["module_list." + std::to_string(linx) + ".Conv2d.weight"], emptywts); + assert(conv1); + conv1->setStrideNd(DimsHW{s, s}); + conv1->setPaddingNd(DimsHW{p, p}); + + IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), "module_list." + std::to_string(linx) + ".BatchNorm2d", 1e-4); + + auto creator = getPluginRegistry()->getPluginCreator("Mish_TRT", "1"); + const PluginFieldCollection* pluginData = creator->getFieldNames(); + IPluginV2 *pluginObj = creator->createPlugin(("mish" + std::to_string(linx)).c_str(), pluginData); + ITensor* inputTensors[] = {bn1->getOutput(0)}; + auto mish = network->addPluginV2(&inputTensors[0], 1, *pluginObj); + return mish; +} \ No newline at end of file diff --git a/scaled-yolov4/gen_wts.py b/scaled-yolov4/gen_wts.py new file mode 100644 index 0000000..448fd40 --- /dev/null +++ b/scaled-yolov4/gen_wts.py @@ -0,0 +1,23 @@ +import struct +import sys +from models.models import * +from utils import * + +model = Darknet('models/yolov4-csp.cfg', (512, 512)) +weights = sys.argv[1] +device = torch_utils.select_device('0') +if weights.endswith('.pt'): # pytorch format + model.load_state_dict(torch.load(weights, map_location=device)['model']) +else: # darknet format + load_darknet_weights(model, weights) + +with open('yolov4_csp.wts', 'w') as f: + f.write('{}\n'.format(len(model.state_dict().keys()))) + for k, v in model.state_dict().items(): + vr = v.reshape(-1).cpu().numpy() + f.write('{} {} '.format(k, len(vr))) + for vv in vr: + f.write(' ') + f.write(struct.pack('>f',float(vv)).hex()) + f.write('\n') + diff --git a/scaled-yolov4/logging.h b/scaled-yolov4/logging.h new file mode 100644 index 0000000..0b1eca3 --- /dev/null +++ b/scaled-yolov4/logging.h @@ -0,0 +1,507 @@ +/* + * Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef TENSORRT_LOGGING_H +#define TENSORRT_LOGGING_H + +#include "NvInferRuntimeCommon.h" +#include +#include +#include +#include +#include +#include +#include + +using Severity = nvinfer1::ILogger::Severity; + +class LogStreamConsumerBuffer : public std::stringbuf +{ +public: + LogStreamConsumerBuffer(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mOutput(stream) + , mPrefix(prefix) + , mShouldLog(shouldLog) + { + } + + LogStreamConsumerBuffer(LogStreamConsumerBuffer&& other) + : mOutput(other.mOutput) + , mPrefix(other.mPrefix) + , mShouldLog(other.mShouldLog) + { + } + + ~LogStreamConsumerBuffer() + { + // std::streambuf::pbase() gives a pointer to the beginning of the buffered part of the output sequence + // std::streambuf::pptr() gives a pointer to the current position of the output sequence + // if the pointer to the beginning is not equal to the pointer to the current position, + // call putOutput() to log the output to the stream + if (pbase() != pptr()) + { + putOutput(); + } + } + + // synchronizes the stream buffer and returns 0 on success + // synchronizing the stream buffer consists of inserting the buffer contents into the stream, + // resetting the buffer and flushing the stream + virtual int sync() + { + putOutput(); + return 0; + } + + void putOutput() + { + if (mShouldLog) + { + // prepend timestamp + std::time_t timestamp = std::time(nullptr); + tm* tm_local = std::localtime(×tamp); + std::cout << "["; + std::cout << std::setw(2) << std::setfill('0') << 1 + tm_local->tm_mon << "/"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_mday << "/"; + std::cout << std::setw(4) << std::setfill('0') << 1900 + tm_local->tm_year << "-"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_hour << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_min << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_sec << "] "; + // std::stringbuf::str() gets the string contents of the buffer + // insert the buffer contents pre-appended by the appropriate prefix into the stream + mOutput << mPrefix << str(); + // set the buffer to empty + str(""); + // flush the stream + mOutput.flush(); + } + } + + void setShouldLog(bool shouldLog) + { + mShouldLog = shouldLog; + } + +private: + std::ostream& mOutput; + std::string mPrefix; + bool mShouldLog; +}; + +//! +//! \class LogStreamConsumerBase +//! \brief Convenience object used to initialize LogStreamConsumerBuffer before std::ostream in LogStreamConsumer +//! +class LogStreamConsumerBase +{ +public: + LogStreamConsumerBase(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mBuffer(stream, prefix, shouldLog) + { + } + +protected: + LogStreamConsumerBuffer mBuffer; +}; + +//! +//! \class LogStreamConsumer +//! \brief Convenience object used to facilitate use of C++ stream syntax when logging messages. +//! Order of base classes is LogStreamConsumerBase and then std::ostream. +//! This is because the LogStreamConsumerBase class is used to initialize the LogStreamConsumerBuffer member field +//! in LogStreamConsumer and then the address of the buffer is passed to std::ostream. +//! This is necessary to prevent the address of an uninitialized buffer from being passed to std::ostream. +//! Please do not change the order of the parent classes. +//! +class LogStreamConsumer : protected LogStreamConsumerBase, public std::ostream +{ +public: + //! \brief Creates a LogStreamConsumer which logs messages with level severity. + //! Reportable severity determines if the messages are severe enough to be logged. + LogStreamConsumer(Severity reportableSeverity, Severity severity) + : LogStreamConsumerBase(severityOstream(severity), severityPrefix(severity), severity <= reportableSeverity) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(severity <= reportableSeverity) + , mSeverity(severity) + { + } + + LogStreamConsumer(LogStreamConsumer&& other) + : LogStreamConsumerBase(severityOstream(other.mSeverity), severityPrefix(other.mSeverity), other.mShouldLog) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(other.mShouldLog) + , mSeverity(other.mSeverity) + { + } + + void setReportableSeverity(Severity reportableSeverity) + { + mShouldLog = mSeverity <= reportableSeverity; + mBuffer.setShouldLog(mShouldLog); + } + +private: + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + static std::string severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + bool mShouldLog; + Severity mSeverity; +}; + +//! \class Logger +//! +//! \brief Class which manages logging of TensorRT tools and samples +//! +//! \details This class provides a common interface for TensorRT tools and samples to log information to the console, +//! and supports logging two types of messages: +//! +//! - Debugging messages with an associated severity (info, warning, error, or internal error/fatal) +//! - Test pass/fail messages +//! +//! The advantage of having all samples use this class for logging as opposed to emitting directly to stdout/stderr is +//! that the logic for controlling the verbosity and formatting of sample output is centralized in one location. +//! +//! In the future, this class could be extended to support dumping test results to a file in some standard format +//! (for example, JUnit XML), and providing additional metadata (e.g. timing the duration of a test run). +//! +//! TODO: For backwards compatibility with existing samples, this class inherits directly from the nvinfer1::ILogger +//! interface, which is problematic since there isn't a clean separation between messages coming from the TensorRT +//! library and messages coming from the sample. +//! +//! In the future (once all samples are updated to use Logger::getTRTLogger() to access the ILogger) we can refactor the +//! class to eliminate the inheritance and instead make the nvinfer1::ILogger implementation a member of the Logger +//! object. + +class Logger : public nvinfer1::ILogger +{ +public: + Logger(Severity severity = Severity::kWARNING) + : mReportableSeverity(severity) + { + } + + //! + //! \enum TestResult + //! \brief Represents the state of a given test + //! + enum class TestResult + { + kRUNNING, //!< The test is running + kPASSED, //!< The test passed + kFAILED, //!< The test failed + kWAIVED //!< The test was waived + }; + + //! + //! \brief Forward-compatible method for retrieving the nvinfer::ILogger associated with this Logger + //! \return The nvinfer1::ILogger associated with this Logger + //! + //! TODO Once all samples are updated to use this method to register the logger with TensorRT, + //! we can eliminate the inheritance of Logger from ILogger + //! + nvinfer1::ILogger& getTRTLogger() + { + return *this; + } + + //! + //! \brief Implementation of the nvinfer1::ILogger::log() virtual method + //! + //! Note samples should not be calling this function directly; it will eventually go away once we eliminate the + //! inheritance from nvinfer1::ILogger + //! + void log(Severity severity, const char* msg) override + { + LogStreamConsumer(mReportableSeverity, severity) << "[TRT] " << std::string(msg) << std::endl; + } + + //! + //! \brief Method for controlling the verbosity of logging output + //! + //! \param severity The logger will only emit messages that have severity of this level or higher. + //! + void setReportableSeverity(Severity severity) + { + mReportableSeverity = severity; + } + + //! + //! \brief Opaque handle that holds logging information for a particular test + //! + //! This object is an opaque handle to information used by the Logger to print test results. + //! The sample must call Logger::defineTest() in order to obtain a TestAtom that can be used + //! with Logger::reportTest{Start,End}(). + //! + class TestAtom + { + public: + TestAtom(TestAtom&&) = default; + + private: + friend class Logger; + + TestAtom(bool started, const std::string& name, const std::string& cmdline) + : mStarted(started) + , mName(name) + , mCmdline(cmdline) + { + } + + bool mStarted; + std::string mName; + std::string mCmdline; + }; + + //! + //! \brief Define a test for logging + //! + //! \param[in] name The name of the test. This should be a string starting with + //! "TensorRT" and containing dot-separated strings containing + //! the characters [A-Za-z0-9_]. + //! For example, "TensorRT.sample_googlenet" + //! \param[in] cmdline The command line used to reproduce the test + // + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + //! + static TestAtom defineTest(const std::string& name, const std::string& cmdline) + { + return TestAtom(false, name, cmdline); + } + + //! + //! \brief A convenience overloaded version of defineTest() that accepts an array of command-line arguments + //! as input + //! + //! \param[in] name The name of the test + //! \param[in] argc The number of command-line arguments + //! \param[in] argv The array of command-line arguments (given as C strings) + //! + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + static TestAtom defineTest(const std::string& name, int argc, char const* const* argv) + { + auto cmdline = genCmdlineString(argc, argv); + return defineTest(name, cmdline); + } + + //! + //! \brief Report that a test has started. + //! + //! \pre reportTestStart() has not been called yet for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has started + //! + static void reportTestStart(TestAtom& testAtom) + { + reportTestResult(testAtom, TestResult::kRUNNING); + assert(!testAtom.mStarted); + testAtom.mStarted = true; + } + + //! + //! \brief Report that a test has ended. + //! + //! \pre reportTestStart() has been called for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has ended + //! \param[in] result The result of the test. Should be one of TestResult::kPASSED, + //! TestResult::kFAILED, TestResult::kWAIVED + //! + static void reportTestEnd(const TestAtom& testAtom, TestResult result) + { + assert(result != TestResult::kRUNNING); + assert(testAtom.mStarted); + reportTestResult(testAtom, result); + } + + static int reportPass(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kPASSED); + return EXIT_SUCCESS; + } + + static int reportFail(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kFAILED); + return EXIT_FAILURE; + } + + static int reportWaive(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kWAIVED); + return EXIT_SUCCESS; + } + + static int reportTest(const TestAtom& testAtom, bool pass) + { + return pass ? reportPass(testAtom) : reportFail(testAtom); + } + + Severity getReportableSeverity() const + { + return mReportableSeverity; + } + +private: + //! + //! \brief returns an appropriate string for prefixing a log message with the given severity + //! + static const char* severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate string for prefixing a test result message with the given result + //! + static const char* testResultString(TestResult result) + { + switch (result) + { + case TestResult::kRUNNING: return "RUNNING"; + case TestResult::kPASSED: return "PASSED"; + case TestResult::kFAILED: return "FAILED"; + case TestResult::kWAIVED: return "WAIVED"; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate output stream (cout or cerr) to use with the given severity + //! + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + //! + //! \brief method that implements logging test results + //! + static void reportTestResult(const TestAtom& testAtom, TestResult result) + { + severityOstream(Severity::kINFO) << "&&&& " << testResultString(result) << " " << testAtom.mName << " # " + << testAtom.mCmdline << std::endl; + } + + //! + //! \brief generate a command line string from the given (argc, argv) values + //! + static std::string genCmdlineString(int argc, char const* const* argv) + { + std::stringstream ss; + for (int i = 0; i < argc; i++) + { + if (i > 0) + { + ss << " "; + } + ss << argv[i]; + } + return ss.str(); + } + + Severity mReportableSeverity; +}; + +namespace +{ + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kVERBOSE +//! +//! Example usage: +//! +//! LOG_VERBOSE(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_VERBOSE(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kVERBOSE); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINFO +//! +//! Example usage: +//! +//! LOG_INFO(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_INFO(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINFO); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kWARNING +//! +//! Example usage: +//! +//! LOG_WARN(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_WARN(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kWARNING); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kERROR +//! +//! Example usage: +//! +//! LOG_ERROR(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_ERROR(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kERROR); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINTERNAL_ERROR +//! ("fatal" severity) +//! +//! Example usage: +//! +//! LOG_FATAL(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_FATAL(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINTERNAL_ERROR); +} + +} // anonymous namespace + +#endif // TENSORRT_LOGGING_H \ No newline at end of file diff --git a/scaled-yolov4/mish.cu b/scaled-yolov4/mish.cu new file mode 100644 index 0000000..d05f609 --- /dev/null +++ b/scaled-yolov4/mish.cu @@ -0,0 +1,196 @@ +#include +#include +#include +#include +#include "mish.h" + +namespace nvinfer1 +{ + MishPlugin::MishPlugin() + { + } + + MishPlugin::~MishPlugin() + { + } + + // create the plugin at runtime from a byte stream + MishPlugin::MishPlugin(const void* data, size_t length) + { + assert(length == sizeof(input_size_)); + input_size_ = *reinterpret_cast(data); + } + + void MishPlugin::serialize(void* buffer) const + { + *reinterpret_cast(buffer) = input_size_; + } + + size_t MishPlugin::getSerializationSize() const + { + return sizeof(input_size_); + } + + int MishPlugin::initialize() + { + return 0; + } + + Dims MishPlugin::getOutputDimensions(int index, const Dims* inputs, int nbInputDims) + { + assert(nbInputDims == 1); + assert(index == 0); + input_size_ = inputs[0].d[0] * inputs[0].d[1] * inputs[0].d[2]; + // Output dimensions + return Dims3(inputs[0].d[0], inputs[0].d[1], inputs[0].d[2]); + } + + // Set plugin namespace + void MishPlugin::setPluginNamespace(const char* pluginNamespace) + { + mPluginNamespace = pluginNamespace; + } + + const char* MishPlugin::getPluginNamespace() const + { + return mPluginNamespace; + } + + // Return the DataType of the plugin output at the requested index + DataType MishPlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const + { + return DataType::kFLOAT; + } + + // Return true if output tensor is broadcast across a batch. + bool MishPlugin::isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const + { + return false; + } + + // Return true if plugin can use input that is broadcast across batch without replication. + bool MishPlugin::canBroadcastInputAcrossBatch(int inputIndex) const + { + return false; + } + + void MishPlugin::configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) + { + } + + // Attach the plugin object to an execution context and grant the plugin the access to some context resource. + void MishPlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) + { + } + + // Detach the plugin object from its execution context. + void MishPlugin::detachFromContext() {} + + const char* MishPlugin::getPluginType() const + { + return "Mish_TRT"; + } + + const char* MishPlugin::getPluginVersion() const + { + return "1"; + } + + void MishPlugin::destroy() + { + delete this; + } + + // Clone the plugin + IPluginV2IOExt* MishPlugin::clone() const + { + MishPlugin *p = new MishPlugin(); + p->input_size_ = input_size_; + p->setPluginNamespace(mPluginNamespace); + return p; + } + + __device__ float tanh_activate_kernel(float x){return (2/(1 + expf(-2*x)) - 1);} + + __device__ float softplus_kernel(float x, float threshold = 20) { + if (x > threshold) return x; // too large + else if (x < -threshold) return expf(x); // too small + return logf(expf(x) + 1); + } + + __global__ void mish_kernel(const float *input, float *output, int num_elem) { + + int idx = threadIdx.x + blockDim.x * blockIdx.x; + if (idx >= num_elem) return; + + //float t = exp(input[idx]); + //if (input[idx] > 20.0) { + // t *= t; + // output[idx] = (t - 1.0) / (t + 1.0); + //} else { + // float tt = t * t; + // output[idx] = (tt + 2.0 * t) / (tt + 2.0 * t + 2.0); + //} + //output[idx] *= input[idx]; + output[idx] = input[idx] * tanh_activate_kernel(softplus_kernel(input[idx])); + } + + void MishPlugin::forwardGpu(const float *const * inputs, float* output, cudaStream_t stream, int batchSize) { + int block_size = thread_count_; + int grid_size = (input_size_ * batchSize + block_size - 1) / block_size; + mish_kernel<<>>(inputs[0], output, input_size_ * batchSize); + } + + int MishPlugin::enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) + { + //assert(batchSize == 1); + //GPU + //CUDA_CHECK(cudaStreamSynchronize(stream)); + forwardGpu((const float *const *)inputs, (float*)outputs[0], stream, batchSize); + return 0; + } + + PluginFieldCollection MishPluginCreator::mFC{}; + std::vector MishPluginCreator::mPluginAttributes; + + MishPluginCreator::MishPluginCreator() + { + mPluginAttributes.clear(); + + mFC.nbFields = mPluginAttributes.size(); + mFC.fields = mPluginAttributes.data(); + } + + const char* MishPluginCreator::getPluginName() const + { + return "Mish_TRT"; + } + + const char* MishPluginCreator::getPluginVersion() const + { + return "1"; + } + + const PluginFieldCollection* MishPluginCreator::getFieldNames() + { + return &mFC; + } + + IPluginV2IOExt* MishPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) + { + MishPlugin* obj = new MishPlugin(); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + + IPluginV2IOExt* MishPluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength) + { + // This object will be deleted when the network is destroyed, which will + // call MishPlugin::destroy() + MishPlugin* obj = new MishPlugin(serialData, serialLength); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + +} + diff --git a/scaled-yolov4/mish.h b/scaled-yolov4/mish.h new file mode 100644 index 0000000..eaed1de --- /dev/null +++ b/scaled-yolov4/mish.h @@ -0,0 +1,108 @@ +#ifndef TRTX_MISH_PLUGIN_H +#define TRTX_MISH_PLUGIN_H + +#include +#include +#include "NvInfer.h" + +namespace nvinfer1 +{ + class MishPlugin: public IPluginV2IOExt + { + public: + explicit MishPlugin(); + MishPlugin(const void* data, size_t length); + + ~MishPlugin(); + + int getNbOutputs() const override + { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override; + + int initialize() override; + + virtual void terminate() override {}; + + virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0;} + + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override; + + virtual size_t getSerializationSize() const override; + + virtual void serialize(void* buffer) const override; + + bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const override { + return inOut[pos].format == TensorFormat::kLINEAR && inOut[pos].type == DataType::kFLOAT; + } + + const char* getPluginType() const override; + + const char* getPluginVersion() const override; + + void destroy() override; + + IPluginV2IOExt* clone() const override; + + void setPluginNamespace(const char* pluginNamespace) override; + + const char* getPluginNamespace() const override; + + DataType getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const override; + + bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const override; + + bool canBroadcastInputAcrossBatch(int inputIndex) const override; + + void attachToContext( + cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) override; + + void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) override; + + void detachFromContext() override; + + int input_size_; + private: + void forwardGpu(const float *const * inputs, float* output, cudaStream_t stream, int batchSize = 1); + int thread_count_ = 256; + const char* mPluginNamespace; + }; + + class MishPluginCreator : public IPluginCreator + { + public: + MishPluginCreator(); + + ~MishPluginCreator() override = default; + + const char* getPluginName() const override; + + const char* getPluginVersion() const override; + + const PluginFieldCollection* getFieldNames() override; + + IPluginV2IOExt* createPlugin(const char* name, const PluginFieldCollection* fc) override; + + IPluginV2IOExt* deserializePlugin(const char* name, const void* serialData, size_t serialLength) override; + + void setPluginNamespace(const char* libNamespace) override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const override + { + return mNamespace.c_str(); + } + + private: + std::string mNamespace; + static PluginFieldCollection mFC; + static std::vector mPluginAttributes; + }; + REGISTER_TENSORRT_PLUGIN(MishPluginCreator); +}; + +#endif // TRTX_MISH_PLUGIN_H \ No newline at end of file diff --git a/scaled-yolov4/utils.h b/scaled-yolov4/utils.h new file mode 100644 index 0000000..e9b3d9c --- /dev/null +++ b/scaled-yolov4/utils.h @@ -0,0 +1,39 @@ +#ifndef __TRT_UTILS_H_ +#define __TRT_UTILS_H_ + +#include +#include +#include +#include + +#ifndef CUDA_CHECK + +#define CUDA_CHECK(callstr) \ + { \ + cudaError_t error_code = callstr; \ + if (error_code != cudaSuccess) { \ + std::cerr << "CUDA error " << error_code << " at " << __FILE__ << ":" << __LINE__; \ + assert(0); \ + } \ + } + +#endif + +namespace Tn +{ + template + void write(char*& buffer, const T& val) + { + *reinterpret_cast(buffer) = val; + buffer += sizeof(T); + } + + template + void read(const char*& buffer, T& val) + { + val = *reinterpret_cast(buffer); + buffer += sizeof(T); + } +} + +#endif \ No newline at end of file diff --git a/scaled-yolov4/yololayer.cu b/scaled-yolov4/yololayer.cu new file mode 100644 index 0000000..e917df0 --- /dev/null +++ b/scaled-yolov4/yololayer.cu @@ -0,0 +1,274 @@ +#include +#include "yololayer.h" +#include "utils.h" + +using namespace Yolo; + +namespace nvinfer1 +{ + YoloLayerPlugin::YoloLayerPlugin() + { + mClassCount = CLASS_NUM; + mYoloKernel.clear(); + mYoloKernel.push_back(yolo1); + mYoloKernel.push_back(yolo2); + mYoloKernel.push_back(yolo3); + + mKernelCount = mYoloKernel.size(); + + CUDA_CHECK(cudaMallocHost(&mAnchor, mKernelCount * sizeof(void*))); + size_t AnchorLen = sizeof(float)* CHECK_COUNT*2; + for(int ii = 0; ii < mKernelCount; ii ++) + { + CUDA_CHECK(cudaMalloc(&mAnchor[ii],AnchorLen)); + const auto& yolo = mYoloKernel[ii]; + CUDA_CHECK(cudaMemcpy(mAnchor[ii], yolo.anchors, AnchorLen, cudaMemcpyHostToDevice)); + } + } + + YoloLayerPlugin::~YoloLayerPlugin() + { + } + + // create the plugin at runtime from a byte stream + YoloLayerPlugin::YoloLayerPlugin(const void* data, size_t length) + { + using namespace Tn; + const char *d = reinterpret_cast(data), *a = d; + read(d, mClassCount); + read(d, mThreadCount); + read(d, mKernelCount); + mYoloKernel.resize(mKernelCount); + auto kernelSize = mKernelCount*sizeof(YoloKernel); + memcpy(mYoloKernel.data(),d,kernelSize); + d += kernelSize; + + CUDA_CHECK(cudaMallocHost(&mAnchor, mKernelCount * sizeof(void*))); + size_t AnchorLen = sizeof(float)* CHECK_COUNT*2; + for(int ii = 0; ii < mKernelCount; ii ++) + { + CUDA_CHECK(cudaMalloc(&mAnchor[ii],AnchorLen)); + const auto& yolo = mYoloKernel[ii]; + CUDA_CHECK(cudaMemcpy(mAnchor[ii], yolo.anchors, AnchorLen, cudaMemcpyHostToDevice)); + } + + assert(d == a + length); + } + + void YoloLayerPlugin::serialize(void* buffer) const + { + using namespace Tn; + char* d = static_cast(buffer), *a = d; + write(d, mClassCount); + write(d, mThreadCount); + write(d, mKernelCount); + auto kernelSize = mKernelCount*sizeof(YoloKernel); + memcpy(d,mYoloKernel.data(),kernelSize); + d += kernelSize; + + assert(d == a + getSerializationSize()); + } + + size_t YoloLayerPlugin::getSerializationSize() const + { + return sizeof(mClassCount) + sizeof(mThreadCount) + sizeof(mKernelCount) + sizeof(Yolo::YoloKernel) * mYoloKernel.size(); + } + + int YoloLayerPlugin::initialize() + { + return 0; + } + + Dims YoloLayerPlugin::getOutputDimensions(int index, const Dims* inputs, int nbInputDims) + { + //output the result to channel + int totalsize = MAX_OUTPUT_BBOX_COUNT * sizeof(Detection) / sizeof(float); + + return Dims3(totalsize + 1, 1, 1); + } + + // Set plugin namespace + void YoloLayerPlugin::setPluginNamespace(const char* pluginNamespace) + { + mPluginNamespace = pluginNamespace; + } + + const char* YoloLayerPlugin::getPluginNamespace() const + { + return mPluginNamespace; + } + + // Return the DataType of the plugin output at the requested index + DataType YoloLayerPlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const + { + return DataType::kFLOAT; + } + + // Return true if output tensor is broadcast across a batch. + bool YoloLayerPlugin::isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const + { + return false; + } + + // Return true if plugin can use input that is broadcast across batch without replication. + bool YoloLayerPlugin::canBroadcastInputAcrossBatch(int inputIndex) const + { + return false; + } + + void YoloLayerPlugin::configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) + { + } + + // Attach the plugin object to an execution context and grant the plugin the access to some context resource. + void YoloLayerPlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) + { + } + + // Detach the plugin object from its execution context. + void YoloLayerPlugin::detachFromContext() {} + + const char* YoloLayerPlugin::getPluginType() const + { + return "YoloLayer_TRT"; + } + + const char* YoloLayerPlugin::getPluginVersion() const + { + return "1"; + } + + void YoloLayerPlugin::destroy() + { + delete this; + } + + // Clone the plugin + IPluginV2IOExt* YoloLayerPlugin::clone() const + { + YoloLayerPlugin *p = new YoloLayerPlugin(); + p->setPluginNamespace(mPluginNamespace); + return p; + } + + __device__ float Logist(float data){ return 1./(1. + exp(-data)); }; + + __global__ void CalDetection(const float *input, float *output, int noElements, + int yoloWidth, int yoloHeight, const float anchors[CHECK_COUNT*2],int classes,int outputElem) { + + int idx = threadIdx.x + blockDim.x * blockIdx.x; + if (idx >= noElements) return; + + int total_grid = yoloWidth * yoloHeight; + int bnIdx = idx / total_grid; + idx = idx - total_grid*bnIdx; + int info_len_i = 5 + classes; + const float* curInput = input + bnIdx * (info_len_i * total_grid * CHECK_COUNT); + + for (int k = 0; k < 3; ++k) { + int class_id = 0; + float max_cls_prob = 0.0; + for (int i = 5; i < info_len_i; ++i) { + float p = Logist(curInput[idx + k * info_len_i * total_grid + i * total_grid]); + if (p > max_cls_prob) { + max_cls_prob = p; + class_id = i - 5; + } + } + float box_prob = Logist(curInput[idx + k * info_len_i * total_grid + 4 * total_grid]); + if (max_cls_prob < IGNORE_THRESH || box_prob < IGNORE_THRESH) continue; + + float *res_count = output + bnIdx*outputElem; + int count = (int)atomicAdd(res_count, 1); + if (count >= MAX_OUTPUT_BBOX_COUNT) return; + char* data = (char * )res_count + sizeof(float) + count*sizeof(Detection); + Detection* det = (Detection*)(data); + + int row = idx / yoloWidth; + int col = idx % yoloWidth; + + //Location + det->bbox[0] = (col + (2 * (Logist(curInput[idx + k * info_len_i * total_grid + 0 * total_grid]))) - 0.5) * INPUT_W / yoloWidth; + det->bbox[1] = (row + (2 * (Logist(curInput[idx + k * info_len_i * total_grid + 1 * total_grid]))) - 0.5) * INPUT_H / yoloHeight; + det->bbox[2] = (powf(2 * (Logist(curInput[idx + k * info_len_i * total_grid + 2 * total_grid])), 2)) * anchors[2*k]; + det->bbox[3] = (powf(2 * (Logist(curInput[idx + k * info_len_i * total_grid + 3 * total_grid])), 2)) * anchors[2*k + 1]; + det->det_confidence = box_prob; + det->class_id = class_id; + det->class_confidence = max_cls_prob; + } + } + + void YoloLayerPlugin::forwardGpu(const float *const * inputs, float* output, cudaStream_t stream, int batchSize) { + + int outputElem = 1 + MAX_OUTPUT_BBOX_COUNT * sizeof(Detection) / sizeof(float); + + for(int idx = 0 ; idx < batchSize; ++idx) { + CUDA_CHECK(cudaMemset(output + idx*outputElem, 0, sizeof(float))); + } + int numElem = 0; + for (unsigned int i = 0;i< mYoloKernel.size();++i) + { + const auto& yolo = mYoloKernel[i]; + numElem = yolo.width*yolo.height*batchSize; + if (numElem < mThreadCount) + mThreadCount = numElem; + CalDetection<<< (yolo.width*yolo.height*batchSize + mThreadCount - 1) / mThreadCount, mThreadCount>>> + (inputs[i],output, numElem, yolo.width, yolo.height, (float *)mAnchor[i], mClassCount ,outputElem); + } + + } + + + int YoloLayerPlugin::enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) + { + //assert(batchSize == 1); + //GPU + //CUDA_CHECK(cudaStreamSynchronize(stream)); + forwardGpu((const float *const *)inputs, (float*)outputs[0], stream, batchSize); + + return 0; + } + + PluginFieldCollection YoloPluginCreator::mFC{}; + std::vector YoloPluginCreator::mPluginAttributes; + + YoloPluginCreator::YoloPluginCreator() + { + mPluginAttributes.clear(); + + mFC.nbFields = mPluginAttributes.size(); + mFC.fields = mPluginAttributes.data(); + } + + const char* YoloPluginCreator::getPluginName() const + { + return "YoloLayer_TRT"; + } + + const char* YoloPluginCreator::getPluginVersion() const + { + return "1"; + } + + const PluginFieldCollection* YoloPluginCreator::getFieldNames() + { + return &mFC; + } + + IPluginV2IOExt* YoloPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) + { + YoloLayerPlugin* obj = new YoloLayerPlugin(); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + + IPluginV2IOExt* YoloPluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength) + { + // This object will be deleted when the network is destroyed, which will + // call MishPlugin::destroy() + YoloLayerPlugin* obj = new YoloLayerPlugin(serialData, serialLength); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + +} \ No newline at end of file diff --git a/scaled-yolov4/yololayer.h b/scaled-yolov4/yololayer.h new file mode 100644 index 0000000..d640404 --- /dev/null +++ b/scaled-yolov4/yololayer.h @@ -0,0 +1,154 @@ +#ifndef _YOLO_LAYER_H +#define _YOLO_LAYER_H + +#include +#include +#include "NvInfer.h" + +namespace Yolo +{ + static constexpr int CHECK_COUNT = 3; + static constexpr float IGNORE_THRESH = 0.1f; + static constexpr int MAX_OUTPUT_BBOX_COUNT = 1000; + static constexpr int CLASS_NUM = 80; + static constexpr int INPUT_H = 512; + static constexpr int INPUT_W = 512; + + struct YoloKernel + { + int width; + int height; + float anchors[CHECK_COUNT*2]; + }; + + static constexpr YoloKernel yolo1 = { + INPUT_W / 8, + INPUT_H / 8, + {12,16, 19,36, 40,28} + }; + static constexpr YoloKernel yolo2 = { + INPUT_W / 16, + INPUT_H / 16, + {36,75, 76,55, 72,146} + }; + static constexpr YoloKernel yolo3 = { + INPUT_W / 32, + INPUT_H / 32, + {142,110, 192,243, 459,401} + }; + + static constexpr int LOCATIONS = 4; + struct alignas(float) Detection{ + //x y w h + float bbox[LOCATIONS]; + float det_confidence; + float class_id; + float class_confidence; + }; +} + + +namespace nvinfer1 +{ + class YoloLayerPlugin: public IPluginV2IOExt + { + public: + explicit YoloLayerPlugin(); + YoloLayerPlugin(const void* data, size_t length); + + ~YoloLayerPlugin(); + + int getNbOutputs() const override + { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override; + + int initialize() override; + + virtual void terminate() override {}; + + virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0;} + + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override; + + virtual size_t getSerializationSize() const override; + + virtual void serialize(void* buffer) const override; + + bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const override { + return inOut[pos].format == TensorFormat::kLINEAR && inOut[pos].type == DataType::kFLOAT; + } + + const char* getPluginType() const override; + + const char* getPluginVersion() const override; + + void destroy() override; + + IPluginV2IOExt* clone() const override; + + void setPluginNamespace(const char* pluginNamespace) override; + + const char* getPluginNamespace() const override; + + DataType getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const override; + + bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const override; + + bool canBroadcastInputAcrossBatch(int inputIndex) const override; + + void attachToContext( + cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) override; + + void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) override; + + void detachFromContext() override; + + private: + void forwardGpu(const float *const * inputs,float * output, cudaStream_t stream, int batchSize = 1); + int mClassCount; + int mKernelCount; + std::vector mYoloKernel; + int mThreadCount = 256; + void** mAnchor; + const char* mPluginNamespace; + }; + + class YoloPluginCreator : public IPluginCreator + { + public: + YoloPluginCreator(); + + ~YoloPluginCreator() override = default; + + const char* getPluginName() const override; + + const char* getPluginVersion() const override; + + const PluginFieldCollection* getFieldNames() override; + + IPluginV2IOExt* createPlugin(const char* name, const PluginFieldCollection* fc) override; + + IPluginV2IOExt* deserializePlugin(const char* name, const void* serialData, size_t serialLength) override; + + void setPluginNamespace(const char* libNamespace) override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const override + { + return mNamespace.c_str(); + } + + private: + std::string mNamespace; + static PluginFieldCollection mFC; + static std::vector mPluginAttributes; + }; + REGISTER_TENSORRT_PLUGIN(YoloPluginCreator); +}; + +#endif diff --git a/scaled-yolov4/yolov4_csp.cpp b/scaled-yolov4/yolov4_csp.cpp new file mode 100644 index 0000000..62014ab --- /dev/null +++ b/scaled-yolov4/yolov4_csp.cpp @@ -0,0 +1,516 @@ +#include +#include +#include + +#include "logging.h" +#include "utils.h" +#include "cuda_runtime_api.h" +#include "common.hpp" + +#define USE_FP16 // comment out this if want to use FP32 +#define DEVICE 0 // GPU id +#define NMS_THRESH 0.4 +#define BBOX_CONF_THRESH 0.5 +#define BATCH_SIZE 1 + +// stuff we know about the network and the input/output blobs +static const int INPUT_H = Yolo::INPUT_H; +static const int INPUT_W = Yolo::INPUT_W; +static const int DETECTION_SIZE = sizeof(Yolo::Detection) / sizeof(float); +static const int OUTPUT_SIZE = Yolo::MAX_OUTPUT_BBOX_COUNT * DETECTION_SIZE + 1; // we assume the yololayer outputs no more than MAX_OUTPUT_BBOX_COUNT boxes that conf >= 0.1 +const char* INPUT_BLOB_NAME = "data"; +const char* OUTPUT_BLOB_NAME = "prob"; + +static Logger gLogger; + + +// Creat the engine using only the API and not any parser. +ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt) { + INetworkDefinition* network = builder -> createNetworkV2(0U); + + // Create input tensor of shape {3, INPUT_H, INPUT_W} with name INPUT_BLOB_NAME + ITensor* data = network -> addInput(INPUT_BLOB_NAME, dt, Dims3{3, INPUT_H, INPUT_W}); + assert(data); + + std::map weightMap = loadWeights("../yolov4_csp.wts"); + Weights emptywts{DataType::kFLOAT, nullptr, 0}; + + // define yolov4 csp layers + auto l0 = convBnMish(network, weightMap, *data, 32, 3, 1, 1, 0); + auto l1 = convBnMish(network, weightMap, *l0 -> getOutput(0), 64, 3, 2, 1, 1); + auto l2 = convBnMish(network, weightMap, *l1 -> getOutput(0), 32, 1, 1, 0, 2); + auto l3 = convBnMish(network, weightMap, *l2 -> getOutput(0), 64, 3, 1, 1, 3); + auto ew4 = network -> addElementWise(*l3 -> getOutput(0), *l1 -> getOutput(0), ElementWiseOperation::kSUM); + auto l5 = convBnMish(network, weightMap, *ew4 -> getOutput(0), 128, 3, 2, 1, 5); + auto l6 = convBnMish(network, weightMap, *l5 -> getOutput(0), 64, 1, 1, 0, 6); + auto l7 = l5; + auto l8 = convBnMish(network, weightMap, *l7 -> getOutput(0), 64, 1, 1, 0, 8); + auto l9 = convBnMish(network, weightMap, *l8 -> getOutput(0), 64, 1, 1, 0, 9); + auto l10 = convBnMish(network, weightMap, *l9 -> getOutput(0), 64, 3, 1, 1, 10); + auto ew11 = network -> addElementWise(*l10 -> getOutput(0), *l8 -> getOutput(0), ElementWiseOperation::kSUM); + auto l12 = convBnMish(network, weightMap, *ew11 -> getOutput(0), 64, 1, 1, 0, 12); + auto l13 = convBnMish(network, weightMap, *l12 -> getOutput(0), 64, 3, 1, 1, 13); + auto ew14 = network -> addElementWise(*l13 -> getOutput(0), *ew11 -> getOutput(0), ElementWiseOperation::kSUM); + auto l15 = convBnMish(network, weightMap, *ew14 -> getOutput(0), 64, 1, 1, 0, 15); + + ITensor* inputTensors16[] = {l15 -> getOutput(0), l6 -> getOutput(0)}; + auto cat16 = network -> addConcatenation(inputTensors16, 2); + + auto l17 = convBnMish(network, weightMap, *cat16 -> getOutput(0), 128, 1, 1, 0, 17); + auto l18 = convBnMish(network, weightMap, *l17 -> getOutput(0), 256, 3, 2, 1, 18); + auto l19 = convBnMish(network, weightMap, *l18 -> getOutput(0), 128, 1, 1, 0, 19); + auto l20 = l18; + auto l21 = convBnMish(network, weightMap, *l20 -> getOutput(0), 128, 1, 1, 0, 21); + auto l22 = convBnMish(network, weightMap, *l21 -> getOutput(0), 128, 1, 1, 0, 22); + auto l23 = convBnMish(network, weightMap, *l22 -> getOutput(0), 128, 3, 1, 1, 23); + auto ew24 = network -> addElementWise(*l23 -> getOutput(0), *l21 -> getOutput(0), ElementWiseOperation::kSUM); + auto l25 = convBnMish(network, weightMap, *ew24 -> getOutput(0), 128, 1, 1, 0, 25); + auto l26 = convBnMish(network, weightMap, *l25 -> getOutput(0), 128, 3, 1, 1, 26); + auto ew27 = network -> addElementWise(*l26 -> getOutput(0), *ew24 -> getOutput(0), ElementWiseOperation::kSUM); + auto l28 = convBnMish(network, weightMap, *ew27 -> getOutput(0), 128, 1, 1, 0, 28); + auto l29 = convBnMish(network, weightMap, *l28 -> getOutput(0), 128, 3, 1, 1, 29); + auto ew30 = network -> addElementWise(*l29 -> getOutput(0), *ew27 -> getOutput(0), ElementWiseOperation::kSUM); + auto l31 = convBnMish(network, weightMap, *ew30 -> getOutput(0), 128, 1, 1, 0, 31); + auto l32 = convBnMish(network, weightMap, *l31 -> getOutput(0), 128, 3, 1, 1, 32); + auto ew33 = network -> addElementWise(*l32 -> getOutput(0), *ew30 -> getOutput(0), ElementWiseOperation::kSUM); + auto l34 = convBnMish(network, weightMap, *ew33 -> getOutput(0), 128, 1, 1, 0, 34); + auto l35 = convBnMish(network, weightMap, *l34 -> getOutput(0), 128, 3, 1, 1, 35); + auto ew36 = network -> addElementWise(*l35 -> getOutput(0), *ew33 -> getOutput(0), ElementWiseOperation::kSUM); + auto l37 = convBnMish(network, weightMap, *ew36 -> getOutput(0), 128, 1, 1, 0, 37); + auto l38 = convBnMish(network, weightMap, *l37 -> getOutput(0), 128, 3, 1, 1, 38); + auto ew39 = network -> addElementWise(*l38 -> getOutput(0), *ew36 -> getOutput(0), ElementWiseOperation::kSUM); + auto l40 = convBnMish(network, weightMap, *ew39 -> getOutput(0), 128, 1, 1, 0, 40); + auto l41 = convBnMish(network, weightMap, *l40 -> getOutput(0), 128, 3, 1, 1, 41); + auto ew42 = network -> addElementWise(*l41 -> getOutput(0), *ew39 -> getOutput(0), ElementWiseOperation::kSUM); + auto l43 = convBnMish(network, weightMap, *ew42 -> getOutput(0), 128, 1, 1, 0, 43); + auto l44 = convBnMish(network, weightMap, *l43 -> getOutput(0), 128, 3, 1, 1, 44); + auto ew45 = network -> addElementWise(*l44 -> getOutput(0), *ew42 -> getOutput(0), ElementWiseOperation::kSUM); + auto l46 = convBnMish(network, weightMap, *ew45 -> getOutput(0), 128, 1, 1, 0, 46); + + ITensor* inputTensors47[] = {l46 -> getOutput(0), l19 -> getOutput(0)}; + auto cat47 = network -> addConcatenation(inputTensors47, 2); + + auto l48 = convBnMish(network, weightMap, *cat47 -> getOutput(0), 256, 1, 1, 0, 48); + auto l49 = convBnMish(network, weightMap, *l48 -> getOutput(0), 512, 3, 2, 1, 49); + auto l50 = convBnMish(network, weightMap, *l49 -> getOutput(0), 256, 1, 1, 0, 50); + auto l51 = l49; + auto l52 = convBnMish(network, weightMap, *l51 -> getOutput(0), 256, 1, 1, 0, 52); + auto l53 = convBnMish(network, weightMap, *l52 -> getOutput(0), 256, 1, 1, 0, 53); + auto l54 = convBnMish(network, weightMap, *l53 -> getOutput(0), 256, 3, 1, 1, 54); + auto ew55 = network -> addElementWise(*l54 -> getOutput(0), *l52 -> getOutput(0), ElementWiseOperation::kSUM); + auto l56 = convBnMish(network, weightMap, *ew55 -> getOutput(0), 256, 1, 1, 0, 56); + auto l57 = convBnMish(network, weightMap, *l56 -> getOutput(0), 256, 3, 1, 1, 57); + auto ew58 = network -> addElementWise(*l57 -> getOutput(0), *ew55 -> getOutput(0), ElementWiseOperation::kSUM); + auto l59 = convBnMish(network, weightMap, *ew58 -> getOutput(0), 256, 1, 1, 0, 59); + auto l60 = convBnMish(network, weightMap, *l59 -> getOutput(0), 256, 3, 1, 1, 60); + auto ew61 = network -> addElementWise(*l60 -> getOutput(0), *ew58 -> getOutput(0), ElementWiseOperation::kSUM); + auto l62 = convBnMish(network, weightMap, *ew61 -> getOutput(0), 256, 1, 1, 0, 62); + auto l63 = convBnMish(network, weightMap, *l62 -> getOutput(0), 256, 3, 1, 1, 63); + auto ew64 = network -> addElementWise(*l63 -> getOutput(0), *ew61 -> getOutput(0), ElementWiseOperation::kSUM); + auto l65 = convBnMish(network, weightMap, *ew64 -> getOutput(0), 256, 1, 1, 0, 65); + auto l66 = convBnMish(network, weightMap, *l65 -> getOutput(0), 256, 3, 1, 1, 66); + auto ew67 = network -> addElementWise(*l66 -> getOutput(0), *ew64 -> getOutput(0), ElementWiseOperation::kSUM); + auto l68 = convBnMish(network, weightMap, *ew67 -> getOutput(0), 256, 1, 1, 0, 68); + auto l69 = convBnMish(network, weightMap, *l68 -> getOutput(0), 256, 3, 1, 1, 69); + auto ew70 = network -> addElementWise(*l69 -> getOutput(0), *ew67 -> getOutput(0), ElementWiseOperation::kSUM); + auto l71 = convBnMish(network, weightMap, *ew70 -> getOutput(0), 256, 1, 1, 0, 71); + auto l72 = convBnMish(network, weightMap, *l71 -> getOutput(0), 256, 3, 1, 1, 72); + auto ew73 = network -> addElementWise(*l72 -> getOutput(0), *ew70 -> getOutput(0), ElementWiseOperation::kSUM); + auto l74 = convBnMish(network, weightMap, *ew73 -> getOutput(0), 256, 1, 1, 0, 74); + auto l75 = convBnMish(network, weightMap, *l74 -> getOutput(0), 256, 3, 1, 1, 75); + auto ew76 = network -> addElementWise(*l75 -> getOutput(0), *ew73 -> getOutput(0), ElementWiseOperation::kSUM); + auto l77 = convBnMish(network, weightMap, *ew76 -> getOutput(0), 256, 1, 1, 0, 77); + + ITensor* inputTensors78[] = {l77 -> getOutput(0), l50 -> getOutput(0)}; + auto cat78 = network -> addConcatenation(inputTensors78, 2); + + auto l79 = convBnMish(network, weightMap, *cat78 -> getOutput(0), 512, 1, 1, 0, 79); + auto l80 = convBnMish(network, weightMap, *l79 -> getOutput(0), 1024, 3, 2, 1, 80); + auto l81 = convBnMish(network, weightMap, *l80 -> getOutput(0), 512, 1, 1, 0, 81); + auto l82 = l80; + auto l83 = convBnMish(network, weightMap, *l82 -> getOutput(0), 512, 1, 1, 0, 83); + auto l84 = convBnMish(network, weightMap, *l83 -> getOutput(0), 512, 1, 1, 0, 84); + auto l85 = convBnMish(network, weightMap, *l84 -> getOutput(0), 512, 3, 1, 1, 85); + auto ew86 = network -> addElementWise(*l85 -> getOutput(0), *l83 -> getOutput(0), ElementWiseOperation::kSUM); + auto l87 = convBnMish(network, weightMap, *ew86 -> getOutput(0), 512, 1, 1, 0, 87); + auto l88 = convBnMish(network, weightMap, *l87 -> getOutput(0), 512, 3, 1, 1, 88); + auto ew89 = network -> addElementWise(*l88 -> getOutput(0), *ew86 -> getOutput(0), ElementWiseOperation::kSUM); + auto l90 = convBnMish(network, weightMap, *ew89 -> getOutput(0), 512, 1, 1, 0, 90); + auto l91 = convBnMish(network, weightMap, *l90 -> getOutput(0), 512, 3, 1, 1, 91); + auto ew92 = network -> addElementWise(*l91 -> getOutput(0), *ew89 -> getOutput(0), ElementWiseOperation::kSUM); + auto l93 = convBnMish(network, weightMap, *ew92 -> getOutput(0), 512, 1, 1, 0, 93); + auto l94 = convBnMish(network, weightMap, *l93 -> getOutput(0), 512, 3, 1, 1, 94); + auto ew95 = network -> addElementWise(*l94 -> getOutput(0), *ew92 -> getOutput(0), ElementWiseOperation::kSUM); + auto l96 = convBnMish(network, weightMap, *ew95 -> getOutput(0), 512, 1, 1, 0, 96); + + ITensor* inputTensors97[] = {l96 -> getOutput(0), l81 -> getOutput(0)}; + + auto cat97 = network -> addConcatenation(inputTensors97, 2); + + auto l98 = convBnMish(network, weightMap, *cat97 -> getOutput(0), 1024, 1, 1, 0, 98); + + // ---- + auto l99 = convBnMish(network, weightMap, *l98 -> getOutput(0), 512, 1, 1, 0, 99); + auto l100 = l98; + auto l101 = convBnMish(network, weightMap, *l100 -> getOutput(0), 512, 1, 1, 0, 101); + auto l102 = convBnMish(network, weightMap, *l101 -> getOutput(0), 512, 3, 1, 1, 102); + auto l103 = convBnMish(network, weightMap, *l102 -> getOutput(0), 512, 1, 1, 0, 103); + + auto pool104 = network -> addPoolingNd(*l103 -> getOutput(0), PoolingType::kMAX, DimsHW{5, 5}); + pool104 -> setPaddingNd(DimsHW{2, 2}); + pool104 -> setStrideNd(DimsHW{1, 1}); + + auto l105 = l103; + + auto pool106 = network -> addPoolingNd(*l105 -> getOutput(0), PoolingType::kMAX, DimsHW{9, 9}); + pool106 -> setPaddingNd(DimsHW{4, 4}); + pool106 -> setStrideNd(DimsHW{1, 1}); + + auto l107 = l103; + + auto pool108 = network -> addPoolingNd(*l107 -> getOutput(0), PoolingType::kMAX, DimsHW{13, 13}); + pool108 -> setPaddingNd(DimsHW{6, 6}); + pool108 -> setStrideNd(DimsHW{1, 1}); + + ITensor* inputTensors109[] = {pool108 -> getOutput(0), pool106 -> getOutput(0), pool104 -> getOutput(0), l103 -> getOutput(0)}; + auto cat109 = network -> addConcatenation(inputTensors109, 4); + + // ---- end spp + + auto l110 = convBnMish(network, weightMap, *cat109 -> getOutput(0), 512, 1, 1, 0, 110); + auto l111 = convBnMish(network, weightMap, *l110 -> getOutput(0), 512, 3, 1, 1, 111); + + ITensor* inputTensors112[] = { l111 -> getOutput(0), l99 -> getOutput(0) }; + auto cat112 = network -> addConcatenation(inputTensors112, 2); + + auto l113 = convBnMish(network, weightMap, *cat112 -> getOutput(0), 512, 1, 1, 0, 113); + auto l114 = convBnMish(network, weightMap, *l113 -> getOutput(0), 256, 1, 1, 0, 114); + + float *deval = reinterpret_cast(malloc(sizeof(float) * 256 * 2 * 2)); + for (int i = 0; i < 256 * 2 * 2; i++) { + deval[i] = 1.0; + } + Weights upsamplewts115{DataType::kFLOAT, deval, 256 * 2 * 2}; + IDeconvolutionLayer* upsample115 = network -> addDeconvolutionNd(*l114 -> getOutput(0), 256, DimsHW{2, 2}, upsamplewts115, emptywts); + assert(upsample115); + upsample115 -> setStrideNd(DimsHW{2, 2}); + upsample115 -> setNbGroups(256); + weightMap["upsample115"] = upsamplewts115; + + auto l116 = l79; + auto l117 = convBnMish(network, weightMap, *l116 -> getOutput(0), 256, 1, 1, 0, 117); + + ITensor* inputTensors118[] = {l117 -> getOutput(0), upsample115 -> getOutput(0)}; + auto cat118 = network -> addConcatenation(inputTensors118, 2); + + auto l119 = convBnMish(network, weightMap, *cat118 -> getOutput(0), 256, 1, 1, 0, 119); + auto l120 = convBnMish(network, weightMap, *l119 -> getOutput(0), 256, 1, 1, 0, 120); + auto l121 = l119; + auto l122 = convBnMish(network, weightMap, *l121 -> getOutput(0), 256, 1, 1, 0, 122); + auto l123 = convBnMish(network, weightMap, *l122 -> getOutput(0), 256, 3, 1, 1, 123); + auto l124 = convBnMish(network, weightMap, *l123 -> getOutput(0), 256, 1, 1, 0, 124); + auto l125 = convBnMish(network, weightMap, *l124 -> getOutput(0), 256, 3, 1, 1, 125); + + ITensor* inputTensors126[] = {l125 -> getOutput(0), l120 -> getOutput(0)}; + auto cat126 = network -> addConcatenation(inputTensors126, 2); + + auto l127 = convBnMish(network, weightMap, *cat126 -> getOutput(0), 256, 1, 1, 0, 127); + auto l128 = convBnMish(network, weightMap, *l127 -> getOutput(0), 128, 1, 1, 0, 128); + + Weights upsamplewts129{DataType::kFLOAT, deval, 128 * 2 * 2}; + IDeconvolutionLayer* upsample129 = network -> addDeconvolutionNd(*l128 -> getOutput(0), 128, DimsHW{2, 2}, upsamplewts129, emptywts); + assert(upsample129); + upsample129 -> setStrideNd(DimsHW{2, 2}); + upsample129 -> setNbGroups(128); + + auto l130 = l48; + auto l131 = convBnMish(network, weightMap, *l130 -> getOutput(0), 128, 1, 1, 0, 131); + + ITensor* inputTensors132[] = {l131 -> getOutput(0), upsample129 -> getOutput(0)}; + auto cat132 = network -> addConcatenation(inputTensors132, 2); + + auto l133 = convBnMish(network, weightMap, *cat132 -> getOutput(0), 128, 1, 1, 0, 133); + auto l134 = convBnMish(network, weightMap, *l133 -> getOutput(0), 128, 1, 1, 0, 134); + auto l135 = l133; + auto l136 = convBnMish(network, weightMap, *l135 -> getOutput(0), 128, 1, 1, 0, 136); + auto l137 = convBnMish(network, weightMap, *l136 -> getOutput(0), 128, 3, 1, 1, 137); + auto l138 = convBnMish(network, weightMap, *l137 -> getOutput(0), 128, 1, 1, 0, 138); + auto l139 = convBnMish(network, weightMap, *l138 -> getOutput(0), 128, 3, 1, 1, 139); + + ITensor* inputTensors140[] = {l139 -> getOutput(0), l134 -> getOutput(0)}; + auto cat140 = network -> addConcatenation(inputTensors140, 2); + + auto l141 = convBnMish(network, weightMap, *cat140 -> getOutput(0), 128, 1, 1, 0, 141); + + // --- + auto l142 = convBnMish(network, weightMap, *l141 -> getOutput(0), 256, 3, 1, 1, 142); + IConvolutionLayer* conv143 = network -> addConvolutionNd(*l142 -> getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{1, 1}, weightMap["module_list.143.Conv2d.weight"], weightMap["module_list.143.Conv2d.bias"]); + assert(conv143); + + // 144 is yolo layer + auto l145 = l141; + auto l146 = convBnMish(network, weightMap, *l145 -> getOutput(0), 256, 3, 2, 1, 146); + + ITensor* inputTensors147[] = {l146 -> getOutput(0), l127 -> getOutput(0)}; + auto cat147 = network -> addConcatenation(inputTensors147, 2); + + auto l148 = convBnMish(network, weightMap, *cat147 -> getOutput(0), 256, 1, 1, 0, 148); + auto l149 = convBnMish(network, weightMap, *l148 -> getOutput(0), 256, 1, 1, 0, 149); + auto l150 = l148; + auto l151 = convBnMish(network, weightMap, *l150 -> getOutput(0), 256, 1, 1, 0, 151); + auto l152 = convBnMish(network, weightMap, *l151 -> getOutput(0), 256, 3, 1, 1, 152); + auto l153 = convBnMish(network, weightMap, *l152 -> getOutput(0), 256, 1, 1, 0, 153); + auto l154 = convBnMish(network, weightMap, *l153 -> getOutput(0), 256, 3, 1, 1, 154); + + ITensor* inputTensors155[] = {l154 -> getOutput(0), l149 -> getOutput(0)}; + auto cat155 = network -> addConcatenation(inputTensors155, 2); + + auto l156 = convBnMish(network, weightMap, *cat155 -> getOutput(0), 256, 1, 1, 0, 156); + auto l157 = convBnMish(network, weightMap, *l156 -> getOutput(0), 512, 3, 1, 1, 157); + IConvolutionLayer* conv158 = network -> addConvolutionNd(*l157 -> getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{1, 1}, weightMap["module_list.158.Conv2d.weight"], weightMap["module_list.158.Conv2d.bias"]); + assert(conv158); + // 159 is yolo layer + + auto l160 = l156; + auto l161 = convBnMish(network, weightMap, *l160 -> getOutput(0), 512, 3, 2, 1, 161); + + ITensor* inputTensors162[] = {l161 -> getOutput(0), l113 -> getOutput(0)}; + auto cat162 = network -> addConcatenation(inputTensors162, 2); + + auto l163 = convBnMish(network, weightMap, *cat162 -> getOutput(0), 512, 1, 1, 0, 163); + auto l164 = convBnMish(network, weightMap, *l163 -> getOutput(0), 512, 1, 1, 0, 164); + auto l165 = l163; + auto l166 = convBnMish(network, weightMap, *l165 -> getOutput(0), 512, 1, 1, 0, 166); + auto l167 = convBnMish(network, weightMap, *l166 -> getOutput(0), 512, 3, 1, 1, 167); + auto l168 = convBnMish(network, weightMap, *l167 -> getOutput(0), 512, 1, 1, 0, 168); + auto l169 = convBnMish(network, weightMap, *l168 -> getOutput(0), 512, 3, 1, 1, 169); + + ITensor* inputTensors170[] = {l169 -> getOutput(0), l164 -> getOutput(0)}; + auto cat170 = network -> addConcatenation(inputTensors170, 2); + + auto l171 = convBnMish(network, weightMap, *cat170 -> getOutput(0), 512, 1, 1, 0, 171); + auto l172 = convBnMish(network, weightMap, *l171 -> getOutput(0), 1024, 3, 1, 1, 172); + + IConvolutionLayer* conv173 = network -> addConvolutionNd(*l172 -> getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{1, 1}, weightMap["module_list.173.Conv2d.weight"], weightMap["module_list.173.Conv2d.bias"]); + assert(conv173); + // 174 is yolo layer + + // add yolo plugin + auto creator = getPluginRegistry() -> getPluginCreator("YoloLayer_TRT", "1"); + const PluginFieldCollection* pluginData = creator -> getFieldNames(); + IPluginV2* pluginObj = creator -> createPlugin("yololayer", pluginData); + ITensor* inputTensorsYolo[] = {conv143 -> getOutput(0), conv158 -> getOutput(0), conv173 -> getOutput(0)}; + auto yolo = network -> addPluginV2(inputTensorsYolo, 3, *pluginObj); + + yolo -> getOutput(0) -> setName(OUTPUT_BLOB_NAME); + network -> markOutput(*yolo -> getOutput(0)); + + // Build engine + builder -> setMaxBatchSize(maxBatchSize); + config -> setMaxWorkspaceSize(16 * (1 << 20)); // 16MB +#ifdef USE_FP16 + config -> setFlag(BuilderFlag::kFP16); +#endif + std::cout << "Building tensorrt engine, please wait for a while..." << std::endl; + ICudaEngine* engine = builder -> buildEngineWithConfig(*network, *config); + std::cout << "Build engine successfully!" << std::endl; + + // Don't need the network any more + network -> destroy(); + + + // Release host memory + for (auto& mem : weightMap) + { + free((void*) (mem.second.values)); + } + + return engine; +} + +void APIToModel(unsigned int maxBatchSize, IHostMemory** modelStream) { + // create builder + IBuilder* builder = createInferBuilder(gLogger); + + // create builder config + IBuilderConfig* config = builder -> createBuilderConfig(); + + // Create model to populate the network, then set the outputs and create an engine + ICudaEngine* engine = createEngine(maxBatchSize, builder, config, DataType::kFLOAT); + assert(engine != nullptr); + + // serialize the trt engine + (*modelStream) = engine -> serialize(); + + // Close everything down + engine -> destroy(); + builder -> destroy(); + config -> destroy(); +} + +void doInference(IExecutionContext& context, float* input, float* output, int batchSize) { + const ICudaEngine& engine = context.getEngine(); + + // Pointers to input and output device buffers to pass to engine. + // Engine requires exactly IEngine::getNbBindings() number of buffers. + assert(engine.getNbBindings() == 2); + void* buffers[2]; + + // In order to bind the buffers, we need to know the names of the input and output tensors. + // Note that indices are guaranteed to be less than IEngine::getNbBindings() + const int inputIndex = engine.getBindingIndex(INPUT_BLOB_NAME); + const int outputIndex = engine.getBindingIndex(OUTPUT_BLOB_NAME); + + // Create GPU buffers on device + CUDA_CHECK(cudaMalloc(&buffers[inputIndex], batchSize * 3 * INPUT_H * INPUT_W * sizeof(float))); + CUDA_CHECK(cudaMalloc(&buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float))); + + // Create stream + cudaStream_t stream; + CUDA_CHECK(cudaStreamCreate(&stream)); + + // DMA input batch data to device, infer on the batch asynchronously, and DMA output back to host + CUDA_CHECK(cudaMemcpyAsync(buffers[inputIndex], input, batchSize * 3 * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream)); + context.enqueue(batchSize, buffers, stream, nullptr); + CUDA_CHECK(cudaMemcpyAsync(output, buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream)); + cudaStreamSynchronize(stream); + + // Release stream and buffers + cudaStreamDestroy(stream); + CUDA_CHECK(cudaFree(buffers[inputIndex])); + CUDA_CHECK(cudaFree(buffers[outputIndex])); +} + +int read_files_in_dir(const char* p_dir_name, std::vector &file_names) { + DIR *p_dir = opendir(p_dir_name); + if (p_dir == nullptr) { + return -1; + } + + struct dirent* p_file = nullptr; + while ((p_file = readdir(p_dir)) != nullptr) { + if (strcmp(p_file -> d_name, ".") != 0 && + strcmp(p_file -> d_name, "..") != 0) { + std::string cur_file_name(p_file->d_name); + file_names.push_back(cur_file_name); + } + } + closedir(p_dir); + return 0; +} + +int main(int argc, char** argv){ + cudaSetDevice(DEVICE); + // create a model using the API directly and serialize it to a stream + char *trtModelStream{nullptr}; + size_t size{0}; + + if (argc == 2 && std::string(argv[1]) == "-s") { + IHostMemory* modelStream{nullptr}; + APIToModel(BATCH_SIZE, &modelStream); + assert(modelStream != nullptr); + std::ofstream p("yolov4csp.engine", std::ios::binary); + if (!p) { + std::cerr << "could not open plan output file" << std::endl; + return -1; + } + p.write(reinterpret_cast(modelStream->data()), modelStream->size()); + modelStream->destroy(); + return 0; + } else if (argc == 3 && std::string(argv[1]) == "-d") { + std::ifstream file("yolov4csp.engine", std::ios::binary); + if (file.good()) { + file.seekg(0, file.end); + size = file.tellg(); + file.seekg(0, file.beg); + trtModelStream = new char[size]; + assert(trtModelStream); + file.read(trtModelStream, size); + file.close(); + } + } else { + std::cerr << "arguments not right!" << std::endl; + std::cerr << "./yolov4 -s // serialize model to plan file" << std::endl; + std::cerr << "./yolov4 -d ../samples // deserialize plan file and run inference" << std::endl; + return -1; + } + + std::vector file_names; + if (read_files_in_dir(argv[2], file_names) < 0) { + std::cout << "read_files_in_dir failed." << std::endl; + return -1; + } + + // prepare input data --------------------------- + static float data[BATCH_SIZE * 3 * INPUT_H * INPUT_W]; + //for (int i = 0; i < 3 * INPUT_H * INPUT_W; i++) + // data[i] = 1.0; + static float prob[BATCH_SIZE * OUTPUT_SIZE]; + IRuntime* runtime = createInferRuntime(gLogger); + assert(runtime != nullptr); + ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size); + assert(engine != nullptr); + IExecutionContext* context = engine->createExecutionContext(); + assert(context != nullptr); + delete[] trtModelStream; + + int fcount = 0; + for (int f = 0; f < (int)file_names.size(); f++) { + fcount++; + if (fcount < BATCH_SIZE && f + 1 != (int)file_names.size()) continue; + for (int b = 0; b < fcount; b++) { + cv::Mat img = cv::imread(std::string(argv[2]) + "/" + file_names[f - fcount + 1 + b]); + if (img.empty()) continue; + cv::Mat pr_img = preprocess_img(img); + for (int i = 0; i < INPUT_H * INPUT_W; i++) { + data[b * 3 * INPUT_H * INPUT_W + i] = pr_img.at(i)[2] / 255.0; + data[b * 3 * INPUT_H * INPUT_W + i + INPUT_H * INPUT_W] = pr_img.at(i)[1] / 255.0; + data[b * 3 * INPUT_H * INPUT_W + i + 2 * INPUT_H * INPUT_W] = pr_img.at(i)[0] / 255.0; + } + } + + // Run inference + auto start = std::chrono::system_clock::now(); + doInference(*context, data, prob, BATCH_SIZE); + auto end = std::chrono::system_clock::now(); + std::cout << std::chrono::duration_cast(end - start).count() << "ms" << std::endl; + std::vector> batch_res(fcount); + for (int b = 0; b < fcount; b++) { + auto& res = batch_res[b]; + nms(res, &prob[b * OUTPUT_SIZE], BBOX_CONF_THRESH, NMS_THRESH); + } + for (int b = 0; b < fcount; b++) { + auto& res = batch_res[b]; + //std::cout << res.size() << std::endl; + cv::Mat img = cv::imread(std::string(argv[2]) + "/" + file_names[f - fcount + 1 + b]); + for (size_t j = 0; j < res.size(); j++) { + float *p = (float*)&res[j]; + for (size_t k = 0; k < 7; k++) { + std::cout << p[k] << ", "; + } + std::cout << std::endl; + cv::Rect r = get_rect(img, res[j].bbox); + cv::rectangle(img, r, cv::Scalar(0x27, 0xC1, 0x36), 2); + cv::putText(img, std::to_string((int)res[j].class_id), cv::Point(r.x, r.y - 1), cv::FONT_HERSHEY_PLAIN, 1.2, cv::Scalar(0xFF, 0xFF, 0xFF), 2); + } + cv::imwrite("_" + file_names[f - fcount + 1 + b], img); + } + fcount = 0; + } + + // Destroy the engine + context->destroy(); + engine->destroy(); + runtime->destroy(); + + //Print histogram of the output distribution + //std::cout << "\nOutput:\n\n"; + //for (unsigned int i = 0; i < OUTPUT_SIZE; i++) + //{ + // std::cout << prob[i] << ", "; + // if (i % 10 == 0) std::cout << i / 10 << std::endl; + //} + //std::cout << std::endl; + + return 0; +} \ No newline at end of file