diff --git a/superpoint/CMakeLists.txt b/superpoint/CMakeLists.txt new file mode 100644 index 0000000..07e9a27 --- /dev/null +++ b/superpoint/CMakeLists.txt @@ -0,0 +1,32 @@ +cmake_minimum_required(VERSION 2.6) + +project(SuperPointNet) + +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 -pthread -Wall -Ofast -Wfatal-errors -D_MWAITXINTRIN_H_INCLUDED") + +find_package(OpenCV) +include_directories(${OpenCV_INCLUDE_DIRS}) + +add_executable(supernet ${PROJECT_SOURCE_DIR}/supernet.cpp ${PROJECT_SOURCE_DIR}/utils.cpp) +target_link_libraries(supernet nvinfer) +target_link_libraries(supernet cudart) +target_link_libraries(supernet ${OpenCV_LIBS}) + +add_definitions(-O2 -pthread) \ No newline at end of file diff --git a/superpoint/README.md b/superpoint/README.md new file mode 100644 index 0000000..f4b9b6b --- /dev/null +++ b/superpoint/README.md @@ -0,0 +1,67 @@ +# SuperPoint + +The PyTorch implementation is from [magicleap/SuperPointPretrainedNetwork.](https://github.com/magicleap/SuperPointPretrainedNetwork) + +The pretrained models are from [magicleap/SuperPointPretrainedNetwork.](https://github.com/magicleap/SuperPointPretrainedNetwork) + + +## Config + +- FP16/FP32 can be selected by the macro `USE_FP16` in supernet.cpp +- GPU id and batch size can be selected by the macro `DEVICE` & `BATCH_SIZE` in supernet.cpp + + +## How to Run +1.Generate .wts file from the baseline pytorch implementation of pretrained model. The following example described how to generate superpoint_v1.wts from pytorch implementation of superpoint_v1. +``` +git clone https://github.com/xiang-wuu/SuperPointPretrainedNetwork +cd SuperPointPretrainedNetwork +git checkout deploy +// copy tensorrtx/superpoint/gen_wts.py to here(SuperPointPretrainedNetwork) +python gen_wts.py +// a file 'superpoint_v1.wts' will be generated. +// before running gen_wts.py python script make sure you cloned private fork and checkout to deploy branch. +``` + +2.Put .wts file into tensorrtx/superpoint, build and run +``` +cd tensorrtx/superpoint +mkdir build +cd build +cmake .. +make +./supernet -s SuperPointPretrainedNetwork/superpoint_v1.wts // serialize model to plan file i.e. 'supernet.engine' +``` + +## Run Demo using SuperPointPretrainedNetwork Python Script +The live demo can be run by inffering TensorRT generated engine file or by the pre-trained pytorch weight file , the `demo_superpoint.py` script is modified to infer automatically by either using TensorRT or PyTorch based on the provided input weight file. +``` +cd SuperPointPretrainedNetwork +python demo_superpoint.py assets/nyu_snippet.mp4 --cuda --weights_path tensorrtx/superpoint/build/supernet.engine +// provide absolute path to supernet.engine as input weight file +python demo_superpoint.py assets/nyu_snippet.mp4 --cuda --weights_path superpoint_v1.pth +// execute above command to infer using pytorch pre-trained weight files instead of tensorrt engine file. +``` + +## Output +As from the below result there is no significant difference in the inferred output! + + + + + + + +
+PyTorch + +TensorRT +
+ + + +
+ +## TODO +- [ ] Optimizing post-processing using custom TensorRT layer. +- [ ] Benchmark validation for speed accuracy tradeoff with [hpatches](https://github.com/hpatches/hpatches-benchmark) dataset diff --git a/superpoint/gen_wts.py b/superpoint/gen_wts.py new file mode 100644 index 0000000..23643d3 --- /dev/null +++ b/superpoint/gen_wts.py @@ -0,0 +1,20 @@ +import torch +import struct +from model import SuperPointNet + +model_name = "superpoint_v1" + +net = SuperPointNet() +net.load_state_dict(torch.load("superpoint_v1.pth")) +net = net.cuda() +net.eval() + +f = open(model_name + ".wts", "w") +f.write("{}\n".format(len(net.state_dict().keys()))) +for k, v in net.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") \ No newline at end of file diff --git a/superpoint/logging.h b/superpoint/logging.h new file mode 100644 index 0000000..fa8f3ca --- /dev/null +++ b/superpoint/logging.h @@ -0,0 +1,517 @@ +/* + * Copyright (c) 2019, 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) + { + } + + ~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) noexcept 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 diff --git a/superpoint/supernet.cpp b/superpoint/supernet.cpp new file mode 100644 index 0000000..0dfd202 --- /dev/null +++ b/superpoint/supernet.cpp @@ -0,0 +1,209 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include "NvInfer.h" +#include "utils.h" +#include "cuda_runtime_api.h" +#include "logging.h" + +//#define USE_FP16 // comment out this if want to use FP32 +#define DEVICE 0 // GPU id +#define BATCH_SIZE 1 // currently, only support BATCH=1 + +// stuff we know about the network and the input/output blobs +static const int INPUT_H = 120; +static const int INPUT_W = 160; +const char *INPUT_BLOB_NAME = "data"; +const char *OUTPUT_BLOB_NAME_1 = "semi"; +const char *OUTPUT_BLOB_NAME_2 = "desc"; + +static Logger gLogger; + +// create the engine using only the API and not any parser. +ICudaEngine *createEngine(IBuilder *builder, IBuilderConfig *config, std::string path, 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{1, INPUT_H, INPUT_W}); + assert(data); + + std::map weightMap = loadWeights(path); + + IConvolutionLayer *conv1a = network->addConvolutionNd(*data, 64, DimsHW{3, 3}, weightMap["conv1a.weight"], weightMap["conv1a.bias"]); + assert(conv1a); + conv1a->setStrideNd(DimsHW{1, 1}); + conv1a->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu1 = network->addActivation(*conv1a->getOutput(0), ActivationType::kRELU); + assert(relu1); + + IConvolutionLayer *conv1b = network->addConvolutionNd(*relu1->getOutput(0), 64, DimsHW{3, 3}, weightMap["conv1b.weight"], weightMap["conv1b.bias"]); + assert(conv1b); + conv1b->setStrideNd(DimsHW{1, 1}); + conv1b->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu2 = network->addActivation(*conv1b->getOutput(0), ActivationType::kRELU); + assert(relu2); + + IPoolingLayer *pool1 = network->addPoolingNd(*relu2->getOutput(0), PoolingType::kMAX, DimsHW{2, 2}); + assert(pool1); + pool1->setStrideNd(DimsHW{2, 2}); + + IConvolutionLayer *conv2a = network->addConvolutionNd(*pool1->getOutput(0), 64, DimsHW{3, 3}, weightMap["conv2a.weight"], weightMap["conv2a.bias"]); + assert(conv2a); + conv2a->setStrideNd(DimsHW{1, 1}); + conv2a->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu3 = network->addActivation(*conv2a->getOutput(0), ActivationType::kRELU); + assert(relu3); + + IConvolutionLayer *conv2b = network->addConvolutionNd(*relu3->getOutput(0), 64, DimsHW{3, 3}, weightMap["conv2b.weight"], weightMap["conv2b.bias"]); + assert(conv2b); + conv2b->setStrideNd(DimsHW{1, 1}); + conv2b->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu4 = network->addActivation(*conv2b->getOutput(0), ActivationType::kRELU); + assert(relu4); + + IPoolingLayer *pool2 = network->addPoolingNd(*relu4->getOutput(0), PoolingType::kMAX, DimsHW{2, 2}); + assert(pool2); + pool2->setStrideNd(DimsHW{2, 2}); + + IConvolutionLayer *conv3a = network->addConvolutionNd(*pool2->getOutput(0), 128, DimsHW{3, 3}, weightMap["conv3a.weight"], weightMap["conv3a.bias"]); + assert(conv3a); + conv3a->setStrideNd(DimsHW{1, 1}); + conv3a->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu44 = network->addActivation(*conv3a->getOutput(0), ActivationType::kRELU); + assert(relu44); + + IConvolutionLayer *conv3b = network->addConvolutionNd(*relu44->getOutput(0), 128, DimsHW{3, 3}, weightMap["conv3b.weight"], weightMap["conv3b.bias"]); + assert(conv3b); + conv3b->setStrideNd(DimsHW{1, 1}); + conv3b->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu5 = network->addActivation(*conv3b->getOutput(0), ActivationType::kRELU); + assert(relu5); + + IPoolingLayer *pool3 = network->addPoolingNd(*relu5->getOutput(0), PoolingType::kMAX, DimsHW{2, 2}); + assert(pool3); + pool3->setStrideNd(DimsHW{2, 2}); + + IConvolutionLayer *conv4a = network->addConvolutionNd(*pool3->getOutput(0), 128, DimsHW{3, 3}, weightMap["conv4a.weight"], weightMap["conv4a.bias"]); + assert(conv4a); + conv4a->setStrideNd(DimsHW{1, 1}); + conv4a->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu6 = network->addActivation(*conv4a->getOutput(0), ActivationType::kRELU); + assert(relu6); + + IConvolutionLayer *conv4b = network->addConvolutionNd(*relu6->getOutput(0), 128, DimsHW{3, 3}, weightMap["conv4b.weight"], weightMap["conv4b.bias"]); + assert(conv4b); + conv4b->setStrideNd(DimsHW{1, 1}); + conv4b->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu7 = network->addActivation(*conv4b->getOutput(0), ActivationType::kRELU); + assert(relu7); + + IConvolutionLayer *convPa = network->addConvolutionNd(*relu7->getOutput(0), 256, DimsHW{3, 3}, weightMap["convPa.weight"], weightMap["convPa.bias"]); + assert(convPa); + convPa->setStrideNd(DimsHW{1, 1}); + convPa->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu8 = network->addActivation(*convPa->getOutput(0), ActivationType::kRELU); + assert(relu8); + + IConvolutionLayer *convPb = network->addConvolutionNd(*relu8->getOutput(0), 65, DimsHW{1, 1}, weightMap["convPb.weight"], weightMap["convPb.bias"]); + assert(convPb); + convPb->setStrideNd(DimsHW{1, 1}); + + IConvolutionLayer *convDa = network->addConvolutionNd(*relu7->getOutput(0), 256, DimsHW{3, 3}, weightMap["convDa.weight"], weightMap["convDa.bias"]); + assert(convDa); + convDa->setStrideNd(DimsHW{1, 1}); + convDa->setPaddingNd(DimsHW{1, 1}); + IActivationLayer *relu9 = network->addActivation(*convDa->getOutput(0), ActivationType::kRELU); + assert(relu9); + + IConvolutionLayer *convDb = network->addConvolutionNd(*relu9->getOutput(0), 256, DimsHW{1, 1}, weightMap["convDb.weight"], weightMap["convDb.bias"]); + assert(convDb); + convDb->setStrideNd(DimsHW{1, 1}); + + convPb->getOutput(0)->setName(OUTPUT_BLOB_NAME_1); + std::cout << "set name out1" << std::endl; + network->markOutput(*convPb->getOutput(0)); + + convDb->getOutput(0)->setName(OUTPUT_BLOB_NAME_2); + std::cout << "set name out2" << std::endl; + network->markOutput(*convDb->getOutput(0)); + + // Build engine + builder->setMaxBatchSize(BATCH_SIZE); + config->setMaxWorkspaceSize(1 << 20); + +#ifdef USE_FP16 + config->setFlag(BuilderFlag::kFP16); +#endif + + ICudaEngine *engine = builder->buildEngineWithConfig(*network, *config); + std::cout << "build out" << 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; +} + +// Creat the engine using only the API and not any parser. + +void APIToModel(std::string path, IHostMemory **modelStream) +{ + // Create builder + IBuilder *builder = createInferBuilder(gLogger); + IBuilderConfig *config = builder->createBuilderConfig(); + + // Create model to populate the network, then set the outputs and create an engine + ICudaEngine *engine = createEngine(builder, config, path, DataType::kFLOAT); + assert(engine != nullptr); + + // Serialize the engine + (*modelStream) = engine->serialize(); + + // Close everything down + engine->destroy(); + builder->destroy(); +} + +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 == 3 && std::string(argv[1]) == "-s") + { + IHostMemory *modelStream{nullptr}; + APIToModel(std::string(argv[2]), &modelStream); + assert(modelStream != nullptr); + std::ofstream p("supernet.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 + { + std::cerr << "arguments not right!" << std::endl; + std::cerr << "./supernet -s // serialize model to plan file" << std::endl; + return -1; + } + + return 0; +} diff --git a/superpoint/utils.cpp b/superpoint/utils.cpp new file mode 100644 index 0000000..2aafccc --- /dev/null +++ b/superpoint/utils.cpp @@ -0,0 +1,91 @@ +#include "utils.h" +#include +#include + +// Load weights from files shared with TensorRT samples. +// 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; +} + +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_dir_name); + // cur_file_name += "/"; + // cur_file_name += p_file->d_name; + std::string cur_file_name(p_file->d_name); + file_names.push_back(cur_file_name); + } + } + + closedir(p_dir); + return 0; +} + +void tokenize(const std::string &str, std::vector &tokens, const std::string &delimiters) +{ + // Skip delimiters at beginning. + std::string::size_type lastPos = str.find_first_not_of(delimiters, 0); + + // Find first non-delimiter. + std::string::size_type pos = str.find_first_of(delimiters, lastPos); + + while (std::string::npos != pos || std::string::npos != lastPos) + { + // Found a token, add it to the vector. + tokens.push_back(str.substr(lastPos, pos - lastPos)); + + // Skip delimiters. + lastPos = str.find_first_not_of(delimiters, pos); + + // Find next non-delimiter. + pos = str.find_first_of(delimiters, lastPos); + } +} diff --git a/superpoint/utils.h b/superpoint/utils.h new file mode 100644 index 0000000..830ef55 --- /dev/null +++ b/superpoint/utils.h @@ -0,0 +1,30 @@ +#pragma once + +#include +#include "NvInfer.h" +#include "cuda_runtime_api.h" +#include "assert.h" +#include +#include +#include +#include +#include + + +using namespace nvinfer1; + +#define CHECK(status) \ + do \ + { \ + auto ret = (status); \ + if (ret != 0) \ + { \ + std::cout << "Cuda failure: " << ret; \ + abort(); \ + } \ + } while (0) + + +int read_files_in_dir(const char *p_dir_name, std::vector &file_names); +std::map loadWeights(const std::string file); +void tokenize(const std::string &str, std::vector &tokens, const std::string &delimiters = ","); \ No newline at end of file