diff --git a/retinaface/CMakeLists.txt b/retinaface/CMakeLists.txt index 355fc32..a2cb331 100644 --- a/retinaface/CMakeLists.txt +++ b/retinaface/CMakeLists.txt @@ -29,8 +29,8 @@ cuda_add_library(decodeplugin SHARED ${PROJECT_SOURCE_DIR}/decode.cu) find_package(OpenCV) include_directories(OpenCV_INCLUDE_DIRS) -add_executable(retina_50 ${PROJECT_SOURCE_DIR}/plugin_factory.cpp ${PROJECT_SOURCE_DIR}/retina_r50.cpp) -target_link_libraries(retina_50 nvinfer nvinfer_plugin) +add_executable(retina_50 ${PROJECT_SOURCE_DIR}/retina_r50.cpp) +target_link_libraries(retina_50 nvinfer) target_link_libraries(retina_50 cudart) target_link_libraries(retina_50 decodeplugin) target_link_libraries(retina_50 ${OpenCV_LIBRARIES}) diff --git a/retinaface/README.md b/retinaface/README.md index 8871e33..381c4ce 100644 --- a/retinaface/README.md +++ b/retinaface/README.md @@ -1,11 +1,10 @@ # RetinaFace -## Notice - -- Tested on TX2/TensorRT4 and GTX1080/TensorRT7 -- The pytorch implementation is [biubug6/Pytorch_Retinaface](https://github.com/biubug6/Pytorch_Retinaface), I forked it into + The pytorch implementation is [biubug6/Pytorch_Retinaface](https://github.com/biubug6/Pytorch_Retinaface), I forked it into [wang-xinyu/Pytorch_Retinaface](https://github.com/biubug6/Pytorch_Retinaface) and add genwts.py +This branch is using TensorRT 7 API, branch [trt4->retinaface](https://github.com/wang-xinyu/tensorrtx/tree/trt4/retinaface) is using TensorRT 4. + ## Run ``` @@ -37,3 +36,13 @@ sudo ./retina_r50 -d // deserialize model file and run inference.

+ +## Config + +- Input shape `INPUT_H`, `INPUT_W` defined in `decode.h` +- FP16/FP32 can be selected by the macro `USE_FP16` in `retina_r50.cpp` +- GPU id can be selected by the macro `DEVICE` in `retina_r50.cpp` + +## More Information + +See the readme in [home page.](https://github.com/wang-xinyu/tensorrtx) diff --git a/retinaface/common.h b/retinaface/common.h deleted file mode 100644 index 3b9c30c..0000000 --- a/retinaface/common.h +++ /dev/null @@ -1,356 +0,0 @@ -#ifndef _TRT_COMMON_H_ -#define _TRT_COMMON_H_ -#include "NvInfer.h" -#include "NvOnnxConfig.h" -#include "NvOnnxParser.h" -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - -using namespace std; - -#define CHECK(status) \ - do \ - { \ - auto ret = (status); \ - if (ret != 0) \ - { \ - std::cout << "Cuda failure: " << ret; \ - abort(); \ - } \ - } while (0) - -constexpr long double operator"" _GB(long double val) { return val * (1 << 30); } -constexpr long double operator"" _MB(long double val) { return val * (1 << 20); } -constexpr long double operator"" _KB(long double val) { return val * (1 << 10); } - -// These is necessary if we want to be able to write 1_GB instead of 1.0_GB. -// Since the return type is signed, -1_GB will work as expected. -constexpr long long int operator"" _GB(long long unsigned int val) { return val * (1 << 30); } -constexpr long long int operator"" _MB(long long unsigned int val) { return val * (1 << 20); } -constexpr long long int operator"" _KB(long long unsigned int val) { return val * (1 << 10); } - -// Logger for TensorRT info/warning/errors -class Logger : public nvinfer1::ILogger -{ -public: - - Logger(): Logger(Severity::kWARNING) {} - - Logger(Severity severity): reportableSeverity(severity) {} - - void log(Severity severity, const char* msg) override - { - // suppress messages with severity enum value greater than the reportable - if (severity > reportableSeverity) return; - - switch (severity) - { - case Severity::kINTERNAL_ERROR: std::cerr << "INTERNAL_ERROR: "; break; - case Severity::kERROR: std::cerr << "ERROR: "; break; - case Severity::kWARNING: std::cerr << "WARNING: "; break; - case Severity::kINFO: std::cerr << "INFO: "; break; - default: std::cerr << "UNKNOWN: "; break; - } - std::cerr << msg << std::endl; - } - - Severity reportableSeverity{Severity::kWARNING}; -}; - -// Locate path to file, given its filename or filepath suffix and possible dirs it might lie in -// Function will also walk back MAX_DEPTH dirs from CWD to check for such a file path -inline std::string locateFile(const std::string& filepathSuffix, const std::vector& directories) -{ - const int MAX_DEPTH{10}; - bool found{false}; - std::string filepath; - - for (auto& dir : directories) - { - filepath = dir + filepathSuffix; - - for (int i = 0; i < MAX_DEPTH && !found; i++) - { - std::ifstream checkFile(filepath); - found = checkFile.is_open(); - if (found) break; - filepath = "../" + filepath; // Try again in parent dir - } - - if (found) - { - break; - } - - filepath.clear(); - } - - if (filepath.empty()) { - std::string directoryList = std::accumulate(directories.begin() + 1, directories.end(), directories.front(), - [](const std::string& a, const std::string& b) { return a + "\n\t" + b; }); - throw std::runtime_error("Could not find " + filepathSuffix + " in data directories:\n\t" + directoryList); - } - return filepath; -} - -inline void readPGMFile(const std::string& fileName, uint8_t* buffer, int inH, int inW) -{ - std::ifstream infile(fileName, std::ifstream::binary); - assert(infile.is_open() && "Attempting to read from a file that is not open."); - std::string magic, h, w, max; - infile >> magic >> h >> w >> max; - infile.seekg(1, infile.cur); - infile.read(reinterpret_cast(buffer), inH * inW); -} - -namespace samples_common -{ - -inline void* safeCudaMalloc(size_t memSize) -{ - void* deviceMem; - CHECK(cudaMalloc(&deviceMem, memSize)); - if (deviceMem == nullptr) - { - std::cerr << "Out of memory" << std::endl; - exit(1); - } - return deviceMem; -} - -inline bool isDebug() -{ - return (std::getenv("TENSORRT_DEBUG") ? true : false); -} - -struct InferDeleter -{ - template - void operator()(T* obj) const - { - if (obj) { - obj->destroy(); - } - } -}; - -template -inline std::shared_ptr infer_object(T* obj) -{ - if (!obj) { - throw std::runtime_error("Failed to create object"); - } - return std::shared_ptr(obj, InferDeleter()); -} - -template -inline std::vector argsort(Iter begin, Iter end, bool reverse = false) -{ - std::vector inds(end - begin); - std::iota(inds.begin(), inds.end(), 0); - if (reverse) { - std::sort(inds.begin(), inds.end(), [&begin](size_t i1, size_t i2) { - return begin[i2] < begin[i1]; - }); - } - else - { - std::sort(inds.begin(), inds.end(), [&begin](size_t i1, size_t i2) { - return begin[i1] < begin[i2]; - }); - } - return inds; -} - -inline bool readReferenceFile(const std::string& fileName, std::vector& refVector) -{ - std::ifstream infile(fileName); - if (!infile.is_open()) { - cout << "ERROR: readReferenceFile: Attempting to read from a file that is not open." << endl; - return false; - } - std::string line; - while (std::getline(infile, line)) { - if (line.empty()) continue; - refVector.push_back(line); - } - infile.close(); - return true; -} - -template -inline std::vector classify(const vector& refVector, const result_vector_t& output, const size_t topK) -{ - auto inds = samples_common::argsort(output.cbegin(), output.cend(), true); - std::vector result; - for (size_t k = 0; k < topK; ++k) { - result.push_back(refVector[inds[k]]); - } - return result; -} - -//...LG returns top K indices, not values. -template -inline vector topK(const vector inp, const size_t k) -{ - vector result; - std::vector inds = samples_common::argsort(inp.cbegin(), inp.cend(), true); - result.assign(inds.begin(), inds.begin()+k); - return result; -} - -template -inline bool readASCIIFile(const string& fileName, const size_t size, vector& out) -{ - std::ifstream infile(fileName); - if (!infile.is_open()) { - cout << "ERROR readASCIIFile: Attempting to read from a file that is not open." << endl; - return false; - } - out.clear(); - out.reserve(size); - out.assign(std::istream_iterator(infile), std::istream_iterator()); - infile.close(); - return true; -} - -template -inline bool writeASCIIFile(const string& fileName, const vector& in) -{ - std::ofstream outfile(fileName); - if (!outfile.is_open()) { - cout << "ERROR: writeASCIIFile: Attempting to write to a file that is not open." << endl; - return false; - } - for (auto fn : in) { - outfile << fn << " "; - } - outfile.close(); - return true; -} - -inline void print_version() -{ -//... This can be only done after statically linking this support into parserONNX.library -#if 0 - std::cout << "Parser built against:" << std::endl; - std::cout << " ONNX IR version: " << nvonnxparser::onnx_ir_version_string(onnx::IR_VERSION) << std::endl; -#endif - std::cout << " TensorRT version: " - << NV_TENSORRT_MAJOR << "." - << NV_TENSORRT_MINOR << "." - << NV_TENSORRT_PATCH << "." - << NV_TENSORRT_BUILD << std::endl; -} - -inline string getFileType(const string& filepath) -{ - return filepath.substr(filepath.find_last_of(".") + 1); -} - -inline string toLower(const string& inp) -{ - string out = inp; - std::transform(out.begin(), out.end(), out.begin(), ::tolower); - return out; -} - -inline unsigned int getElementSize(nvinfer1::DataType t) -{ - switch (t) - { - case nvinfer1::DataType::kINT32: return 4; - case nvinfer1::DataType::kFLOAT: return 4; - case nvinfer1::DataType::kHALF: return 2; - case nvinfer1::DataType::kINT8: return 1; - } - throw std::runtime_error("Invalid DataType."); - return 0; -} - -inline int64_t volume(const nvinfer1::Dims& d) -{ - return std::accumulate(d.d, d.d + d.nbDims, 1, std::multiplies()); -} - -// Struct to maintain command-line arguments. -struct Args -{ - bool runInInt8 = false; -}; - -// Populates the Args struct with the provided command-line parameters. -inline void parseArgs(Args& args, int argc, char* argv[]) -{ - if (argc >= 1) - { - for (int i = 1; i < argc; ++i) - { - if (!strcmp(argv[i], "--int8")) args.runInInt8 = true; - } - } -} - -template -struct PPM -{ - std::string magic, fileName; - int h, w, max; - uint8_t buffer[C * H * W]; -}; - -struct BBox -{ - float x1, y1, x2, y2; -}; - -template -inline void writePPMFileWithBBox(const std::string& filename, PPM& ppm, const BBox& bbox) -{ - std::ofstream outfile("./" + filename, std::ofstream::binary); - assert(!outfile.fail()); - outfile << "P6" << "\n" << ppm.w << " " << ppm.h << "\n" << ppm.max << "\n"; - auto round = [](float x) -> int { return int(std::floor(x + 0.5f)); }; - const int x1 = std::min(std::max(0, round(int(bbox.x1))), W - 1); - const int x2 = std::min(std::max(0, round(int(bbox.x2))), W - 1); - const int y1 = std::min(std::max(0, round(int(bbox.y1))), H - 1); - const int y2 = std::min(std::max(0, round(int(bbox.y2))), H - 1); - for (int x = x1; x <= x2; ++x) - { - // bbox top border - ppm.buffer[(y1 * ppm.w + x) * 3] = 255; - ppm.buffer[(y1 * ppm.w + x) * 3 + 1] = 0; - ppm.buffer[(y1 * ppm.w + x) * 3 + 2] = 0; - // bbox bottom border - ppm.buffer[(y2 * ppm.w + x) * 3] = 255; - ppm.buffer[(y2 * ppm.w + x) * 3 + 1] = 0; - ppm.buffer[(y2 * ppm.w + x) * 3 + 2] = 0; - } - for (int y = y1; y <= y2; ++y) - { - // bbox left border - ppm.buffer[(y * ppm.w + x1) * 3] = 255; - ppm.buffer[(y * ppm.w + x1) * 3 + 1] = 0; - ppm.buffer[(y * ppm.w + x1) * 3 + 2] = 0; - // bbox right border - ppm.buffer[(y * ppm.w + x2) * 3] = 255; - ppm.buffer[(y * ppm.w + x2) * 3 + 1] = 0; - ppm.buffer[(y * ppm.w + x2) * 3 + 2] = 0; - } - outfile.write(reinterpret_cast(ppm.buffer), ppm.w * ppm.h * 3); -} - -} // namespace samples_common - -#endif // _TRT_COMMON_H_ diff --git a/retinaface/decode.cu b/retinaface/decode.cu index f04c06b..27da62b 100644 --- a/retinaface/decode.cu +++ b/retinaface/decode.cu @@ -3,24 +3,24 @@ namespace nvinfer1 { - DecodePlugin::DecodePlugin(const int cudaThread):thread_count_(cudaThread) + DecodePlugin::DecodePlugin() { } - + DecodePlugin::~DecodePlugin() { } - + // create the plugin at runtime from a byte stream DecodePlugin::DecodePlugin(const void* data, size_t length) { } - void DecodePlugin::serialize(void* buffer) + void DecodePlugin::serialize(void* buffer) const { } - - size_t DecodePlugin::getSerializationSize() + + size_t DecodePlugin::getSerializationSize() const { return 0; } @@ -29,7 +29,7 @@ namespace nvinfer1 { return 0; } - + Dims DecodePlugin::getOutputDimensions(int index, const Dims* inputs, int nbInputDims) { //output the result to channel @@ -41,10 +41,74 @@ namespace nvinfer1 return Dims3(totalCount + 1, 1, 1); } - __device__ float Logist(float data){ return 1./(1. + exp(-data)); }; + // Set plugin namespace + void DecodePlugin::setPluginNamespace(const char* pluginNamespace) + { + mPluginNamespace = pluginNamespace; + } + + const char* DecodePlugin::getPluginNamespace() const + { + return mPluginNamespace; + } + + // Return the DataType of the plugin output at the requested index + DataType DecodePlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const + { + return DataType::kFLOAT; + } + + // Return true if output tensor is broadcast across a batch. + bool DecodePlugin::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 DecodePlugin::canBroadcastInputAcrossBatch(int inputIndex) const + { + return false; + } + + void DecodePlugin::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 DecodePlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) + { + } + + // Detach the plugin object from its execution context. + void DecodePlugin::detachFromContext() {} + + const char* DecodePlugin::getPluginType() const + { + return "Decode_TRT"; + } + + const char* DecodePlugin::getPluginVersion() const + { + return "1"; + } + + void DecodePlugin::destroy() + { + delete this; + } + + // Clone the plugin + IPluginV2IOExt* DecodePlugin::clone() const + { + DecodePlugin *p = new DecodePlugin(); + p->setPluginNamespace(mPluginNamespace); + return p; + } + + __device__ float Logist(float data){ return 1./(1. + expf(-data)); }; __global__ void CalDetection(const float *input, float *output, int num_elem, int step, int anchor) { - + int idx = threadIdx.x + blockDim.x * blockIdx.x; if (idx >= num_elem) return; @@ -59,7 +123,7 @@ namespace nvinfer1 for (int k = 0; k < 2; ++k) { float conf1 = cls_reg[idx + k * num_elem * 2]; float conf2 = cls_reg[idx + k * num_elem * 2 + num_elem]; - conf2 = exp(conf2) / (exp(conf1) + exp(conf2)); + conf2 = expf(conf2) / (expf(conf1) + expf(conf2)); if (conf2 <= 0.02) continue; float *res_count = output; @@ -76,8 +140,8 @@ namespace nvinfer1 //Location det->bbox[0] = prior[0] + bbox_reg[idx + k * num_elem * 4] * 0.1 * prior[2]; det->bbox[1] = prior[1] + bbox_reg[idx + k * num_elem * 4 + num_elem] * 0.1 * prior[3]; - det->bbox[2] = prior[2] * exp(bbox_reg[idx + k * num_elem * 4 + num_elem * 2] * 0.2); - det->bbox[3] = prior[3] * exp(bbox_reg[idx + k * num_elem * 4 + num_elem * 3] * 0.2); + det->bbox[2] = prior[2] * expf(bbox_reg[idx + k * num_elem * 4 + num_elem * 2] * 0.2); + det->bbox[3] = prior[3] * expf(bbox_reg[idx + k * num_elem * 4 + num_elem * 3] * 0.2); det->bbox[0] -= det->bbox[2] / 2; det->bbox[1] -= det->bbox[3] / 2; det->bbox[2] += det->bbox[0]; @@ -95,7 +159,7 @@ namespace nvinfer1 } } } - + void DecodePlugin::forwardGpu(const float *const * inputs, float * output, cudaStream_t stream, int batchSize) { int num_elem = 0; @@ -124,4 +188,46 @@ namespace nvinfer1 return 0; }; + PluginFieldCollection DecodePluginCreator::mFC{}; + std::vector DecodePluginCreator::mPluginAttributes; + + DecodePluginCreator::DecodePluginCreator() + { + mPluginAttributes.clear(); + + mFC.nbFields = mPluginAttributes.size(); + mFC.fields = mPluginAttributes.data(); + } + + const char* DecodePluginCreator::getPluginName() const + { + return "Decode_TRT"; + } + + const char* DecodePluginCreator::getPluginVersion() const + { + return "1"; + } + + const PluginFieldCollection* DecodePluginCreator::getFieldNames() + { + return &mFC; + } + + IPluginV2IOExt* DecodePluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) + { + DecodePlugin* obj = new DecodePlugin(); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + + IPluginV2IOExt* DecodePluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength) + { + // This object will be deleted when the network is destroyed, which will + // call PReluPlugin::destroy() + DecodePlugin* obj = new DecodePlugin(serialData, serialLength); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + } diff --git a/retinaface/decode.h b/retinaface/decode.h index 86e4a71..d69ddf7 100644 --- a/retinaface/decode.h +++ b/retinaface/decode.h @@ -1,6 +1,8 @@ #ifndef _DECODE_CU_H #define _DECODE_CU_H +#include +#include #include "NvInfer.h" namespace decodeplugin @@ -14,46 +16,102 @@ namespace decodeplugin static const int INPUT_W = 1600; } - namespace nvinfer1 { - class DecodePlugin: public IPluginExt + class DecodePlugin: public IPluginV2IOExt { - public: - explicit DecodePlugin(const int cudaThread = 256); - DecodePlugin(const void* data, size_t length); + public: + DecodePlugin(); + DecodePlugin(const void* data, size_t length); - ~DecodePlugin(); + ~DecodePlugin(); - int getNbOutputs() const override - { - return 1; - } + int getNbOutputs() const override + { + return 1; + } - Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override; + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override; - bool supportsFormat(DataType type, PluginFormat format) const override { - return type == DataType::kFLOAT && format == PluginFormat::kNCHW; - } + int initialize() override; - void configureWithFormat(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, DataType type, PluginFormat format, int maxBatchSize) override {}; + virtual void terminate() override {}; - int initialize() override; + virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0;} - virtual void terminate() override {}; + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override; - virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0;} + virtual size_t getSerializationSize() const override; - virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override; + virtual void serialize(void* buffer) const override; - virtual size_t getSerializationSize() 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; + } - virtual void serialize(void* buffer) override; + const char* getPluginType() const override; - void forwardGpu(const float *const * inputs,float * output, cudaStream_t stream,int batchSize = 1); + const char* getPluginVersion() const override; - private: - int thread_count_ = 256; + 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 DecodePluginCreator : public IPluginCreator + { + public: + DecodePluginCreator(); + + ~DecodePluginCreator() 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; }; }; diff --git a/retinaface/logging.h b/retinaface/logging.h new file mode 100644 index 0000000..602b69f --- /dev/null +++ b/retinaface/logging.h @@ -0,0 +1,503 @@ +/* + * 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) 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/retinaface/plugin_factory.cpp b/retinaface/plugin_factory.cpp deleted file mode 100644 index 0ee35a7..0000000 --- a/retinaface/plugin_factory.cpp +++ /dev/null @@ -1,17 +0,0 @@ -#include "plugin_factory.h" -#include "NvInferPlugin.h" -#include "decode.h" -#include "common.h" - -using namespace nvinfer1; -using nvinfer1::PluginFactory; - -IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialData, size_t serialLength) { - IPlugin *plugin = nullptr; - if (strstr(layerName, "leaky") != NULL) { - plugin = plugin::createPReLUPlugin(serialData, serialLength); - } else if (strstr(layerName, "decode") != NULL) { - plugin = new DecodePlugin(serialData, serialLength); - } - return plugin; -} diff --git a/retinaface/plugin_factory.h b/retinaface/plugin_factory.h deleted file mode 100644 index 0be0225..0000000 --- a/retinaface/plugin_factory.h +++ /dev/null @@ -1,12 +0,0 @@ -#ifndef MY_PLUGIN_FACTORY_H -#define MY_PLUGIN_FACTORY_H -#include - -namespace nvinfer1 { -class PluginFactory : public IPluginFactory { - public: - IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength) override; -}; - -} -#endif diff --git a/retinaface/retina_r50.cpp b/retinaface/retina_r50.cpp index 148b81c..1a122e1 100644 --- a/retinaface/retina_r50.cpp +++ b/retinaface/retina_r50.cpp @@ -1,18 +1,27 @@ -#include "NvInfer.h" -#include "NvInferPlugin.h" -#include "cuda_runtime_api.h" -#include "common.h" #include #include #include #include #include #include -#include "plugin_factory.h" -#include "decode.h" #include +#include "NvInfer.h" +#include "cuda_runtime_api.h" +#include "decode.h" +#include "logging.h" -//#define USE_FP16 // comment out this if want to use FP32 +#define CHECK(status) \ + do\ + {\ + auto ret = (status);\ + if (ret != 0)\ + {\ + std::cerr << "Cuda failure: " << ret << std::endl;\ + abort();\ + }\ + } while (0) + +#define USE_FP16 // comment out this if want to use FP32 #define DEVICE 0 // GPU id // stuff we know about the network and the input/output blobs @@ -24,6 +33,7 @@ const char* OUTPUT_BLOB_NAME = "prob"; using namespace nvinfer1; static Logger gLogger; +REGISTER_TENSORRT_PLUGIN(DecodePluginCreator); cv::Mat preprocess_img(cv::Mat& img) { int w, h, x, y; @@ -75,10 +85,10 @@ cv::Rect get_rect_adapt_landmark(cv::Mat& img, float bbox[4], float lmk[10]) { float iou(float lbox[4], float rbox[4]) { float interBox[] = { - max(lbox[0], rbox[0]), //left - min(lbox[2], rbox[2]), //right - max(lbox[1], rbox[1]), //top - min(lbox[3], rbox[3]), //bottom + std::max(lbox[0], rbox[0]), //left + std::min(lbox[2], rbox[2]), //right + std::max(lbox[1], rbox[1]), //top + std::min(lbox[3], rbox[3]), //bottom }; if(interBox[2] > interBox[3] || interBox[0] > interBox[1]) @@ -170,7 +180,6 @@ IScaleLayer* addBatchNorm2d(INetworkDefinition *network, std::map(malloc(sizeof(float) * len)); for (int i = 0; i < len; i++) { @@ -201,7 +210,7 @@ IScaleLayer* addBatchNorm2d(INetworkDefinition *network, std::map& weightMap, ITensor& input, int inch, int outch, int stride, std::string lname) { Weights emptywts{DataType::kFLOAT, nullptr, 0}; - IConvolutionLayer* conv1 = network->addConvolution(input, outch, DimsHW{1, 1}, weightMap[lname + "conv1.weight"], emptywts); + IConvolutionLayer* conv1 = network->addConvolutionNd(input, outch, DimsHW{1, 1}, weightMap[lname + "conv1.weight"], emptywts); assert(conv1); IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), lname + "bn1", 1e-5); @@ -209,26 +218,26 @@ IActivationLayer* bottleneck(INetworkDefinition *network, std::mapaddActivation(*bn1->getOutput(0), ActivationType::kRELU); assert(relu1); - IConvolutionLayer* conv2 = network->addConvolution(*relu1->getOutput(0), outch, DimsHW{3, 3}, weightMap[lname + "conv2.weight"], emptywts); + IConvolutionLayer* conv2 = network->addConvolutionNd(*relu1->getOutput(0), outch, DimsHW{3, 3}, weightMap[lname + "conv2.weight"], emptywts); assert(conv2); - conv2->setStride(DimsHW{stride, stride}); - conv2->setPadding(DimsHW{1, 1}); + conv2->setStrideNd(DimsHW{stride, stride}); + conv2->setPaddingNd(DimsHW{1, 1}); IScaleLayer* bn2 = addBatchNorm2d(network, weightMap, *conv2->getOutput(0), lname + "bn2", 1e-5); IActivationLayer* relu2 = network->addActivation(*bn2->getOutput(0), ActivationType::kRELU); assert(relu2); - IConvolutionLayer* conv3 = network->addConvolution(*relu2->getOutput(0), outch * 4, DimsHW{1, 1}, weightMap[lname + "conv3.weight"], emptywts); + IConvolutionLayer* conv3 = network->addConvolutionNd(*relu2->getOutput(0), outch * 4, DimsHW{1, 1}, weightMap[lname + "conv3.weight"], emptywts); assert(conv3); IScaleLayer* bn3 = addBatchNorm2d(network, weightMap, *conv3->getOutput(0), lname + "bn3", 1e-5); IElementWiseLayer* ew1; if (stride != 1 || inch != outch * 4) { - IConvolutionLayer* conv4 = network->addConvolution(input, outch * 4, DimsHW{1, 1}, weightMap[lname + "downsample.0.weight"], emptywts); + IConvolutionLayer* conv4 = network->addConvolutionNd(input, outch * 4, DimsHW{1, 1}, weightMap[lname + "downsample.0.weight"], emptywts); assert(conv4); - conv4->setStride(DimsHW{stride, stride}); + conv4->setStrideNd(DimsHW{stride, stride}); IScaleLayer* bn4 = addBatchNorm2d(network, weightMap, *conv4->getOutput(0), lname + "downsample.1", 1e-5); ew1 = network->addElementWise(*bn4->getOutput(0), *bn3->getOutput(0), ElementWiseOperation::kSUM); @@ -243,10 +252,10 @@ IActivationLayer* bottleneck(INetworkDefinition *network, std::map& weightMap, ITensor& input, int outch, int kernelsize, int stride, int padding, bool userelu, std::string lname) { Weights emptywts{DataType::kFLOAT, nullptr, 0}; - IConvolutionLayer* conv1 = network->addConvolution(input, outch, DimsHW{kernelsize, kernelsize}, getWeights(weightMap, lname + ".0.weight"), emptywts); + IConvolutionLayer* conv1 = network->addConvolutionNd(input, outch, DimsHW{kernelsize, kernelsize}, getWeights(weightMap, lname + ".0.weight"), emptywts); assert(conv1); - conv1->setStride(DimsHW{stride, stride}); - conv1->setPadding(DimsHW{padding, padding}); + conv1->setStrideNd(DimsHW{stride, stride}); + conv1->setPaddingNd(DimsHW{padding, padding}); IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), lname + ".1", 1e-5); @@ -272,10 +281,10 @@ IActivationLayer* ssh(INetworkDefinition *network, std::mapcreateNetwork(); +ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt) { + INetworkDefinition* network = builder->createNetworkV2(0U); - // Create input tensor of shape { 1, 1, 32, 32 } with name INPUT_BLOB_NAME + // Create input tensor with name INPUT_BLOB_NAME ITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims3{3, INPUT_H, INPUT_W}); assert(data); @@ -283,10 +292,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType Weights emptywts{DataType::kFLOAT, nullptr, 0}; // ------------- backbone resnet50 --------------- - IConvolutionLayer* conv1 = network->addConvolution(*data, 64, DimsHW{7, 7}, weightMap["body.conv1.weight"], emptywts); + IConvolutionLayer* conv1 = network->addConvolutionNd(*data, 64, DimsHW{7, 7}, weightMap["body.conv1.weight"], emptywts); assert(conv1); - conv1->setStride(DimsHW{2, 2}); - conv1->setPadding(DimsHW{3, 3}); + conv1->setStrideNd(DimsHW{2, 2}); + conv1->setPaddingNd(DimsHW{3, 3}); IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), "body.bn1", 1e-5); @@ -295,10 +304,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType assert(relu1); // Add max pooling layer with stride of 2x2 and kernel size of 2x2. - IPoolingLayer* pool1 = network->addPooling(*relu1->getOutput(0), PoolingType::kMAX, DimsHW{3, 3}); + IPoolingLayer* pool1 = network->addPoolingNd(*relu1->getOutput(0), PoolingType::kMAX, DimsHW{3, 3}); assert(pool1); - pool1->setStride(DimsHW{2, 2}); - pool1->setPadding(DimsHW{1, 1}); + pool1->setStrideNd(DimsHW{2, 2}); + pool1->setPaddingNd(DimsHW{1, 1}); IActivationLayer* x = bottleneck(network, weightMap, *pool1->getOutput(0), 64, 64, 1, "body.layer1.0."); x = bottleneck(network, weightMap, *x->getOutput(0), 256, 64, 1, "body.layer1.1."); @@ -333,18 +342,18 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType deval[i] = 1.0; } Weights deconvwts{DataType::kFLOAT, deval, 256 * 2 * 2}; - IDeconvolutionLayer* up3 = network->addDeconvolution(*output3->getOutput(0), 256, DimsHW{2, 2}, deconvwts, emptywts); + IDeconvolutionLayer* up3 = network->addDeconvolutionNd(*output3->getOutput(0), 256, DimsHW{2, 2}, deconvwts, emptywts); assert(up3); - up3->setStride(DimsHW{2, 2}); + up3->setStrideNd(DimsHW{2, 2}); up3->setNbGroups(256); weightMap["up3"] = deconvwts; output2 = network->addElementWise(*output2->getOutput(0), *up3->getOutput(0), ElementWiseOperation::kSUM); output2 = conv_bn_relu(network, weightMap, *output2->getOutput(0), 256, 3, 1, 1, true, "fpn.merge2"); - IDeconvolutionLayer* up2 = network->addDeconvolution(*output2->getOutput(0), 256, DimsHW{2, 2}, deconvwts, emptywts); + IDeconvolutionLayer* up2 = network->addDeconvolutionNd(*output2->getOutput(0), 256, DimsHW{2, 2}, deconvwts, emptywts); assert(up2); - up2->setStride(DimsHW{2, 2}); + up2->setStrideNd(DimsHW{2, 2}); up2->setNbGroups(256); output1 = network->addElementWise(*output1->getOutput(0), *up2->getOutput(0), ElementWiseOperation::kSUM); output1 = conv_bn_relu(network, weightMap, *output1->getOutput(0), 256, 3, 1, 1, true, "fpn.merge1"); @@ -355,17 +364,17 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType auto ssh3 = ssh(network, weightMap, *output3->getOutput(0), "ssh3"); // ------------- Head --------------- - auto bbox_head1 = network->addConvolution(*ssh1->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.0.conv1x1.weight"], weightMap["BboxHead.0.conv1x1.bias"]); - auto bbox_head2 = network->addConvolution(*ssh2->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.1.conv1x1.weight"], weightMap["BboxHead.1.conv1x1.bias"]); - auto bbox_head3 = network->addConvolution(*ssh3->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.2.conv1x1.weight"], weightMap["BboxHead.2.conv1x1.bias"]); + auto bbox_head1 = network->addConvolutionNd(*ssh1->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.0.conv1x1.weight"], weightMap["BboxHead.0.conv1x1.bias"]); + auto bbox_head2 = network->addConvolutionNd(*ssh2->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.1.conv1x1.weight"], weightMap["BboxHead.1.conv1x1.bias"]); + auto bbox_head3 = network->addConvolutionNd(*ssh3->getOutput(0), 2 * 4, DimsHW{1, 1}, weightMap["BboxHead.2.conv1x1.weight"], weightMap["BboxHead.2.conv1x1.bias"]); - auto cls_head1 = network->addConvolution(*ssh1->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.0.conv1x1.weight"], weightMap["ClassHead.0.conv1x1.bias"]); - auto cls_head2 = network->addConvolution(*ssh2->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.1.conv1x1.weight"], weightMap["ClassHead.1.conv1x1.bias"]); - auto cls_head3 = network->addConvolution(*ssh3->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.2.conv1x1.weight"], weightMap["ClassHead.2.conv1x1.bias"]); + auto cls_head1 = network->addConvolutionNd(*ssh1->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.0.conv1x1.weight"], weightMap["ClassHead.0.conv1x1.bias"]); + auto cls_head2 = network->addConvolutionNd(*ssh2->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.1.conv1x1.weight"], weightMap["ClassHead.1.conv1x1.bias"]); + auto cls_head3 = network->addConvolutionNd(*ssh3->getOutput(0), 2 * 2, DimsHW{1, 1}, weightMap["ClassHead.2.conv1x1.weight"], weightMap["ClassHead.2.conv1x1.bias"]); - auto lmk_head1 = network->addConvolution(*ssh1->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.0.conv1x1.weight"], weightMap["LandmarkHead.0.conv1x1.bias"]); - auto lmk_head2 = network->addConvolution(*ssh2->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.1.conv1x1.weight"], weightMap["LandmarkHead.1.conv1x1.bias"]); - auto lmk_head3 = network->addConvolution(*ssh3->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.2.conv1x1.weight"], weightMap["LandmarkHead.2.conv1x1.bias"]); + auto lmk_head1 = network->addConvolutionNd(*ssh1->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.0.conv1x1.weight"], weightMap["LandmarkHead.0.conv1x1.bias"]); + auto lmk_head2 = network->addConvolutionNd(*ssh2->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.1.conv1x1.weight"], weightMap["LandmarkHead.1.conv1x1.bias"]); + auto lmk_head3 = network->addConvolutionNd(*ssh3->getOutput(0), 2 * 10, DimsHW{1, 1}, weightMap["LandmarkHead.2.conv1x1.weight"], weightMap["LandmarkHead.2.conv1x1.bias"]); // ------------- Decode bbox, conf, landmark --------------- ITensor* inputTensors1[] = {bbox_head1->getOutput(0), cls_head1->getOutput(0), lmk_head1->getOutput(0)}; @@ -374,24 +383,26 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType auto cat2 = network->addConcatenation(inputTensors2, 3); ITensor* inputTensors3[] = {bbox_head3->getOutput(0), cls_head3->getOutput(0), lmk_head3->getOutput(0)}; auto cat3 = network->addConcatenation(inputTensors3, 3); - auto decode = new DecodePlugin(); + + auto creator = getPluginRegistry()->getPluginCreator("Decode_TRT", "1"); + PluginFieldCollection pfc; + IPluginV2 *pluginObj = creator->createPlugin("decode", &pfc); ITensor* inputTensors[] = {cat1->getOutput(0), cat2->getOutput(0), cat3->getOutput(0)}; - auto decodelayer = network->addPlugin(inputTensors, 3, *decode); + auto decodelayer = network->addPluginV2(inputTensors, 3, *pluginObj); assert(decodelayer); - decodelayer->setName("decode"); decodelayer->getOutput(0)->setName(OUTPUT_BLOB_NAME); - std::cout << "set name out, start building trt engine..." << std::endl; network->markOutput(*decodelayer->getOutput(0)); // Build engine builder->setMaxBatchSize(maxBatchSize); - builder->setMaxWorkspaceSize(1 << 20); + config->setMaxWorkspaceSize(1 << 20); #ifdef USE_FP16 - builder->setFp16Mode(true); + config->setFlag(BuilderFlag::kFP16); #endif - ICudaEngine* engine = builder->buildCudaEngine(*network); - std::cout << "build out" << std::endl; + std::cout << "Building 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(); @@ -409,9 +420,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, DataType void APIToModel(unsigned int maxBatchSize, 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(maxBatchSize, builder, DataType::kFLOAT); + ICudaEngine* engine = createEngine(maxBatchSize, builder, config, DataType::kFLOAT); assert(engine != nullptr); // Serialize the engine @@ -473,9 +485,8 @@ int main(int argc, char** argv) { APIToModel(1, &modelStream); assert(modelStream != nullptr); - std::ofstream p("retina_r50.engine"); - if (!p) - { + std::ofstream p("retina_r50.engine", std::ios::binary); + if (!p) { std::cerr << "could not open plan output file" << std::endl; return -1; } @@ -511,10 +522,9 @@ int main(int argc, char** argv) { data[i + 2 * INPUT_H * INPUT_W] = pr_img.at(i)[2] - 123.0; } - PluginFactory pf; IRuntime* runtime = createInferRuntime(gLogger); assert(runtime != nullptr); - ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size, &pf); + ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size); //ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size, nullptr); assert(engine != nullptr); IExecutionContext* context = engine->createExecutionContext();