diff --git a/.gitignore b/.gitignore index 11da3e2..1e5457c 100644 --- a/.gitignore +++ b/.gitignore @@ -1,2 +1,3 @@ */build */*.wts +*/*.ppm diff --git a/README.md b/README.md index 9cee18c..0c2dea7 100644 --- a/README.md +++ b/README.md @@ -8,9 +8,10 @@ I wrote this project to get familiar with tensorrt API, and also to share and le All the models are implemented in pytorch first, and export a weights file xxx.wts, and then use tensorrt to load weights, define network and do inference. Some pytorch implementations can be found in my repo [Pytorchx](https://github.com/wang-xinyu/pytorchx), the remaining are from polular open-source pytorch implementations. -# News +## News - `22 May 2020`. A new branch [trt4](https://github.com/wang-xinyu/tensorrtx/tree/trt4) created, which is using TensorRT 4 API. Now the master branch is using TensorRT 7 API. But only `yolov4` has been migrated to TensorRT 7 API for now. The rest will be migrated soon. And a tutorial for `migarating from TensorRT 4 to 7` provided. +- `28 May 2020`. arcface LResNet50E-IR model from [deepinsight/insightface](https://github.com/deepinsight/insightface) implemented. We got 333fps on GTX1080. ## Tutorials @@ -47,6 +48,7 @@ Following models are implemented. |[yolov3-spp](./yolov3-spp)| darknet-53, weights from [ultralytics/yolov3](https://github.com/ultralytics/yolov3) | |[yolov4](./yolov4)| CSPDarknet53, weights from [AlexeyAB/darknet](https://github.com/AlexeyAB/darknet#pre-trained-models), pytorch implementation from [ultralytics/yolov3](https://github.com/ultralytics/yolov3) | |[retinaface](./retinaface)| resnet-50, weights from [biubug6/Pytorch_Retinaface](https://github.com/biubug6/Pytorch_Retinaface) | +|[arcface](./arcface)| LResNet50E-IR, weights from [deepinsight/insightface](https://github.com/deepinsight/insightface) | ## Tricky Operations @@ -68,6 +70,7 @@ Some tricky operations encountered in these models, already solved, but might ha |hsigmoid| hard sigmoid is implemented as a plugin, hsigmoid and hswish are used in mobilenetv3 | |retinaface output decode| implement a plugin to decode bbox, confidence and landmarks, see retinaface. | |mish| mish activation is implemented as a plugin, mish is used in yolov4 | +|prelu| mxnet's prelu activation with trainable gamma is implemented as a plugin, used in arcface | ## Speed Benchmark @@ -80,6 +83,7 @@ Some tricky operations encountered in these models, already solved, but might ha | YOLOv4(CSPDarknet53) | Xeon E5-2620/GTX1080 | 8 | FP16 | 608x608 | 41.3 | | RetinaFace(resnet50) | TX2 | 1 | FP16 | 384x640 | 15 | | RetinaFace(resnet50) | Xeon E5-2620/GTX1080 | 1 | FP32 | 928x1600 | 15 | +| ArcFace(LResNet50E-IR) | Xeon E5-2620/GTX1080 | 1 | FP32 | 112x112 | 333 | Detection net FPS test including inference and nms time, excluding image preprocess time. diff --git a/arcface/CMakeLists.txt b/arcface/CMakeLists.txt new file mode 100644 index 0000000..db0e030 --- /dev/null +++ b/arcface/CMakeLists.txt @@ -0,0 +1,41 @@ +cmake_minimum_required(VERSION 2.6) + +project(arcface) + +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) + +set(CUDA_NVCC_PLAGS ${CUDA_NVCC_PLAGS};-std=c++11;-g;-G;-gencode;arch=compute_30;code=sm_30) + +include_directories(${PROJECT_SOURCE_DIR}/include) +if (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64") + message("embed_platform on") + include_directories(/usr/local/cuda/targets/aarch64-linux/include) + link_directories(/usr/local/cuda/targets/aarch64-linux/lib) +else() + message("embed_platform off") + include_directories(/usr/local/cuda/include) + link_directories(/usr/local/cuda/lib64) +endif() + + +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}/prelu.cu) + +find_package(OpenCV) +include_directories(OpenCV_INCLUDE_DIRS) + +add_executable(arcface-r50 ${PROJECT_SOURCE_DIR}/arcface-r50.cpp) +target_link_libraries(arcface-r50 nvinfer) +target_link_libraries(arcface-r50 cudart) +target_link_libraries(arcface-r50 myplugins) +target_link_libraries(arcface-r50 ${OpenCV_LIBS}) + +add_definitions(-O2 -pthread) + diff --git a/arcface/README.md b/arcface/README.md new file mode 100644 index 0000000..902114b --- /dev/null +++ b/arcface/README.md @@ -0,0 +1,47 @@ +# arcface + +The mxnet implementation is from [deepinsight/insightface.](https://github.com/deepinsight/insightface) + +The pretrained model is [LResNet50E-IR,ArcFace@ms1m-refine-v1.](https://github.com/deepinsight/insightface/wiki/Model-Zoo#32-lresnet50e-irarcfacems1m-refine-v1) + +The two images used in this project are joey0.ppm and joey1.ppm, download them from [Google Drive.](https://drive.google.com/drive/folders/1ctqpkRCRKyBZRCNwo9Uq4eUoMRLtFq1e) + +

+![joey](https://user-images.githubusercontent.com/15235574/83122953-f45f8d80-a106-11ea-84b0-4f6ff91b5924.jpg) +

+ +## Run + +``` +1. generate arcface-r50.wts from mxnet implementation with LResNet50E-IR,ArcFace@ms1m-refine-v1 pretrained model + +git clone https://github.com/deepinsight/insightface +cd insightface/deploy +// copy tensorrtx/arcface/gen_wts.py to here(insightface/deploy) +// download model-r50-am-lfw.zip and unzip here(insightface/deploy) +python gen_wts.py +// a file 'arcface-r50.wts' will be generated. +// the master branch of insightface should work, if not, you can checkout 94ad870abb3203d6f31b049b70dd080dc8f33fca + +2. put arcface-r50.wts into tensorrtx/arcface, build and run + +cd tensorrtx/arcface +// download joey0.ppm and joey1.ppm, and put here(tensorrtx/arcface) +mkdir build +cd build +cmake .. +make +sudo ./arcface-r50 -s // serialize model to plan file i.e. 'arcface-r50.engine' +sudo ./arcface-r50 -d // deserialize plan file and run inference + +3. check the output log, latency and similarity score. +``` + +## Config + +- FP16/FP32 can be selected by the macro `USE_FP16` in arcface-r50.cpp +- GPU id can be selected by the macro `DEVICE` in arcface-r50.cpp + +## More Information + +See the readme in [home page.](https://github.com/wang-xinyu/tensorrtx) diff --git a/arcface/arcface-r50.cpp b/arcface/arcface-r50.cpp new file mode 100644 index 0000000..ac0079c --- /dev/null +++ b/arcface/arcface-r50.cpp @@ -0,0 +1,403 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include "NvInfer.h" +#include "cuda_runtime_api.h" +#include "logging.h" +#include "prelu.h" + +#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 +#define BATCH_SIZE 1 // currently, only support BATCH=1 + +using namespace nvinfer1; + +// stuff we know about the network and the input/output blobs +static const int INPUT_H = 112; +static const int INPUT_W = 112; +static const int OUTPUT_SIZE = 512; +const char* INPUT_BLOB_NAME = "data"; +const char* OUTPUT_BLOB_NAME = "prob"; +static Logger gLogger; +REGISTER_TENSORRT_PLUGIN(PReluPluginCreator); + +// 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 + "_gamma"].values; + float *beta = (float*)weightMap[lname + "_beta"].values; + float *mean = (float*)weightMap[lname + "_moving_mean"].values; + float *var = (float*)weightMap[lname + "_moving_var"].values; + int len = weightMap[lname + "_moving_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* addPRelu(INetworkDefinition *network, std::map& weightMap, ITensor& input, std::string lname) { + auto creator = getPluginRegistry()->getPluginCreator("PRelu_TRT", "1"); + PluginFieldCollection pfc; + PluginField pf("gamma", weightMap[lname + "_gamma"].values, PluginFieldType::kFLOAT32, weightMap[lname + "_gamma"].count); + pfc.nbFields = 1; + pfc.fields = &pf; + IPluginV2 *pluginObj = creator->createPlugin(lname.c_str(), &pfc); + ITensor* inputTensors[] = {&input}; + auto prelu = network->addPluginV2(&inputTensors[0], 1, *pluginObj); + assert(prelu); + return prelu; +} + +ILayer* resUnit(INetworkDefinition *network, std::map& weightMap, ITensor& input, int num_filters, int s, bool dim_match, std::string lname) { + Weights emptywts{DataType::kFLOAT, nullptr, 0}; + auto bn1 = addBatchNorm2d(network, weightMap, input, lname + "_bn1", 2e-5); + IConvolutionLayer* conv1 = network->addConvolutionNd(*bn1->getOutput(0), num_filters, DimsHW{3, 3}, weightMap[lname + "_conv1_weight"], emptywts); + assert(conv1); + conv1->setPaddingNd(DimsHW{1, 1}); + auto bn2 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), lname + "_bn2", 2e-5); + auto act1 = addPRelu(network, weightMap, *bn2->getOutput(0), lname + "_relu1"); + IConvolutionLayer* conv2 = network->addConvolutionNd(*act1->getOutput(0), num_filters, DimsHW{3, 3}, weightMap[lname + "_conv2_weight"], emptywts); + assert(conv2); + conv2->setStrideNd(DimsHW{s, s}); + conv2->setPaddingNd(DimsHW{1, 1}); + auto bn3 = addBatchNorm2d(network, weightMap, *conv2->getOutput(0), lname + "_bn3", 2e-5); + + IElementWiseLayer* ew1; + if (dim_match) { + ew1 = network->addElementWise(input, *bn3->getOutput(0), ElementWiseOperation::kSUM); + } else { + IConvolutionLayer* conv1sc = network->addConvolutionNd(input, num_filters, DimsHW{1, 1}, weightMap[lname + "_conv1sc_weight"], emptywts); + assert(conv1sc); + conv1sc->setStrideNd(DimsHW{s, s}); + auto bn1sc = addBatchNorm2d(network, weightMap, *conv1sc->getOutput(0), lname + "_sc", 2e-5); + ew1 = network->addElementWise(*bn1sc->getOutput(0), *bn3->getOutput(0), ElementWiseOperation::kSUM); + } + assert(ew1); + return ew1; +} + +// 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("../arcface-r50.wts"); + Weights emptywts{DataType::kFLOAT, nullptr, 0}; + + IConvolutionLayer* conv0 = network->addConvolutionNd(*data, 64, DimsHW{3, 3}, weightMap["conv0_weight"], emptywts); + assert(conv0); + conv0->setPaddingNd(DimsHW{1, 1}); + auto bn0 = addBatchNorm2d(network, weightMap, *conv0->getOutput(0), "bn0", 2e-5); + auto relu0 = addPRelu(network, weightMap, *bn0->getOutput(0), "relu0"); + + auto s1u1 = resUnit(network, weightMap, *relu0->getOutput(0), 64, 2, false, "stage1_unit1"); + auto s1u2 = resUnit(network, weightMap, *s1u1->getOutput(0), 64, 1, true, "stage1_unit2"); + auto s1u3 = resUnit(network, weightMap, *s1u2->getOutput(0), 64, 1, true, "stage1_unit3"); + + auto s2u1 = resUnit(network, weightMap, *s1u3->getOutput(0), 128, 2, false, "stage2_unit1"); + auto s2u2 = resUnit(network, weightMap, *s2u1->getOutput(0), 128, 1, true, "stage2_unit2"); + auto s2u3 = resUnit(network, weightMap, *s2u2->getOutput(0), 128, 1, true, "stage2_unit3"); + auto s2u4 = resUnit(network, weightMap, *s2u3->getOutput(0), 128, 1, true, "stage2_unit4"); + + auto s3u1 = resUnit(network, weightMap, *s2u4->getOutput(0), 256, 2, false, "stage3_unit1"); + auto s3u2 = resUnit(network, weightMap, *s3u1->getOutput(0), 256, 1, true, "stage3_unit2"); + auto s3u3 = resUnit(network, weightMap, *s3u2->getOutput(0), 256, 1, true, "stage3_unit3"); + auto s3u4 = resUnit(network, weightMap, *s3u3->getOutput(0), 256, 1, true, "stage3_unit4"); + auto s3u5 = resUnit(network, weightMap, *s3u4->getOutput(0), 256, 1, true, "stage3_unit5"); + auto s3u6 = resUnit(network, weightMap, *s3u5->getOutput(0), 256, 1, true, "stage3_unit6"); + auto s3u7 = resUnit(network, weightMap, *s3u6->getOutput(0), 256, 1, true, "stage3_unit7"); + auto s3u8 = resUnit(network, weightMap, *s3u7->getOutput(0), 256, 1, true, "stage3_unit8"); + auto s3u9 = resUnit(network, weightMap, *s3u8->getOutput(0), 256, 1, true, "stage3_unit9"); + auto s3u10 = resUnit(network, weightMap, *s3u9->getOutput(0), 256, 1, true, "stage3_unit10"); + auto s3u11 = resUnit(network, weightMap, *s3u10->getOutput(0), 256, 1, true, "stage3_unit11"); + auto s3u12 = resUnit(network, weightMap, *s3u11->getOutput(0), 256, 1, true, "stage3_unit12"); + auto s3u13 = resUnit(network, weightMap, *s3u12->getOutput(0), 256, 1, true, "stage3_unit13"); + auto s3u14 = resUnit(network, weightMap, *s3u13->getOutput(0), 256, 1, true, "stage3_unit14"); + + auto s4u1 = resUnit(network, weightMap, *s3u14->getOutput(0), 512, 2, false, "stage4_unit1"); + auto s4u2 = resUnit(network, weightMap, *s4u1->getOutput(0), 512, 1, true, "stage4_unit2"); + auto s4u3 = resUnit(network, weightMap, *s4u2->getOutput(0), 512, 1, true, "stage4_unit3"); + + auto bn1 = addBatchNorm2d(network, weightMap, *s4u3->getOutput(0), "bn1", 2e-5); + IFullyConnectedLayer* fc1 = network->addFullyConnected(*bn1->getOutput(0), 512, weightMap["pre_fc1_weight"], weightMap["pre_fc1_bias"]); + assert(fc1); + auto bn2 = addBatchNorm2d(network, weightMap, *fc1->getOutput(0), "fc1", 2e-5); + + bn2->getOutput(0)->setName(OUTPUT_BLOB_NAME); + network->markOutput(*bn2->getOutput(0)); + + // Build engine + builder->setMaxBatchSize(maxBatchSize); + config->setMaxWorkspaceSize(16 * (1 << 20)); // 16MB +#ifdef USE_FP16 + config->setFlag(BuilderFlag::kFP16); +#endif + 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(); + + // 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); + 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 engine + (*modelStream) = engine->serialize(); + + // Close everything down + engine->destroy(); + builder->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 + CHECK(cudaMalloc(&buffers[inputIndex], batchSize * 3 * INPUT_H * INPUT_W * sizeof(float))); + CHECK(cudaMalloc(&buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float))); + + // Create stream + cudaStream_t stream; + CHECK(cudaStreamCreate(&stream)); + + // DMA input batch data to device, infer on the batch asynchronously, and DMA output back to host + CHECK(cudaMemcpyAsync(buffers[inputIndex], input, batchSize * 3 * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream)); + context.enqueue(batchSize, buffers, stream, nullptr); + CHECK(cudaMemcpyAsync(output, buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream)); + cudaStreamSynchronize(stream); + + // Release stream and buffers + cudaStreamDestroy(stream); + CHECK(cudaFree(buffers[inputIndex])); + 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_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; +} + +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("arcface-r50.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 == 2 && std::string(argv[1]) == "-d") { + std::ifstream file("arcface-r50.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 << "./arcface-r50 -s // serialize model to plan file" << std::endl; + std::cerr << "./arcface-r50 -d // deserialize plan file and run inference" << 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; + + cv::Mat img = cv::imread("../joey0.ppm"); + for (int i = 0; i < INPUT_H * INPUT_W; i++) { + data[i] = ((float)img.at(i)[0] - 127.5) * 0.0078125; + data[i + INPUT_H * INPUT_W] = ((float)img.at(i)[1] - 127.5) * 0.0078125; + data[i + 2 * INPUT_H * INPUT_W] = ((float)img.at(i)[2] - 127.5) * 0.0078125; + } + + // 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; + + cv::Mat out(512, 1, CV_32FC1, prob); + cv::Mat out_norm; + cv::normalize(out, out_norm); + + img = cv::imread("../joey1.ppm"); + for (int i = 0; i < INPUT_H * INPUT_W; i++) { + data[i] = ((float)img.at(i)[0] - 127.5) * 0.0078125; + data[i + INPUT_H * INPUT_W] = ((float)img.at(i)[1] - 127.5) * 0.0078125; + data[i + 2 * INPUT_H * INPUT_W] = ((float)img.at(i)[2] - 127.5) * 0.0078125; + } + + // Run inference + start = std::chrono::system_clock::now(); + doInference(*context, data, prob, BATCH_SIZE); + end = std::chrono::system_clock::now(); + std::cout << std::chrono::duration_cast(end - start).count() << "ms" << std::endl; + + cv::Mat out1(1, 512, CV_32FC1, prob); + cv::Mat out_norm1; + cv::normalize(out1, out_norm1); + + cv::Mat res = out_norm1 * out_norm; + + std::cout << "similarity score: " << *(float*)res.data << std::endl; + + // 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 << p_out_norm[i] << ", "; + // if (i % 10 == 0) std::cout << i / 10 << std::endl; + //} + //std::cout << std::endl; + + return 0; +} diff --git a/arcface/logging.h b/arcface/logging.h new file mode 100644 index 0000000..602b69f --- /dev/null +++ b/arcface/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/arcface/prelu.cu b/arcface/prelu.cu new file mode 100644 index 0000000..0fea0e2 --- /dev/null +++ b/arcface/prelu.cu @@ -0,0 +1,210 @@ +#include +#include +#include +#include +#include "prelu.h" + +namespace nvinfer1 +{ + PReluPlugin::PReluPlugin(const std::vector& gamma) : gamma_(gamma) + { + } + + PReluPlugin::~PReluPlugin() + { + } + + // create the plugin at runtime from a byte stream + PReluPlugin::PReluPlugin(const void* data, size_t length) + { + char *p = (char*)data; + input_size_ = reinterpret_cast(p)[0]; + p += sizeof(int); + gamma_.assign((float*)p, (float*)p + (length - sizeof(int)) / sizeof(float)); + } + + void PReluPlugin::serialize(void* buffer) const + { + *reinterpret_cast(buffer) = input_size_; + char *p = reinterpret_cast(buffer); + p += sizeof(int); + memcpy(p, gamma_.data(), gamma_.size() * sizeof(float)); + } + + size_t PReluPlugin::getSerializationSize() const + { + return sizeof(input_size_) + gamma_.size() * sizeof(float); + } + + int PReluPlugin::initialize() + { + return 0; + } + + Dims PReluPlugin::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 PReluPlugin::setPluginNamespace(const char* pluginNamespace) + { + mPluginNamespace = pluginNamespace; + } + + const char* PReluPlugin::getPluginNamespace() const + { + return mPluginNamespace; + } + + // Return the DataType of the plugin output at the requested index + DataType PReluPlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const + { + return DataType::kFLOAT; + } + + // Return true if output tensor is broadcast across a batch. + bool PReluPlugin::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 PReluPlugin::canBroadcastInputAcrossBatch(int inputIndex) const + { + return false; + } + + void PReluPlugin::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 PReluPlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) + { + } + + // Detach the plugin object from its execution context. + void PReluPlugin::detachFromContext() {} + + const char* PReluPlugin::getPluginType() const + { + return "PRelu_TRT"; + } + + const char* PReluPlugin::getPluginVersion() const + { + return "1"; + } + + void PReluPlugin::destroy() + { + delete this; + } + + // Clone the plugin + IPluginV2IOExt* PReluPlugin::clone() const + { + PReluPlugin *p = new PReluPlugin(gamma_); + p->input_size_ = input_size_; + p->setPluginNamespace(mPluginNamespace); + return p; + } + + __global__ void prelu_kernel(const float *input, float *output, int num_elem, int input_size, int fm_size, const float* gamma) { + + int idx = threadIdx.x + blockDim.x * blockIdx.x; + if (idx >= num_elem) return; + + if (input[idx] >= 0.0f) { + output[idx] = input[idx]; + return; + } + int c = (idx % input_size) / fm_size; + output[idx] = input[idx] * gamma[c]; + } + + void PReluPlugin::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; + void *dev_gamma; + assert(cudaMalloc(&dev_gamma, sizeof(float) * gamma_.size()) == cudaSuccess); + assert(cudaMemcpy(dev_gamma, gamma_.data(), sizeof(float) * gamma_.size(), cudaMemcpyHostToDevice) == cudaSuccess); + prelu_kernel<<>>(inputs[0], output, input_size_ * batchSize, input_size_, input_size_ / gamma_.size(), (const float*)dev_gamma); + assert(cudaFree(dev_gamma) == cudaSuccess); + } + + int PReluPlugin::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 PReluPluginCreator::mFC{}; + std::vector PReluPluginCreator::mPluginAttributes; + + PReluPluginCreator::PReluPluginCreator() + { + mPluginAttributes.emplace_back(PluginField("gamma", nullptr, PluginFieldType::kFLOAT32, 1)); + + mFC.nbFields = mPluginAttributes.size(); + mFC.fields = mPluginAttributes.data(); + } + + const char* PReluPluginCreator::getPluginName() const + { + return "PRelu_TRT"; + } + + const char* PReluPluginCreator::getPluginVersion() const + { + return "1"; + } + + const PluginFieldCollection* PReluPluginCreator::getFieldNames() + { + return &mFC; + } + + IPluginV2IOExt* PReluPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) + { + std::vector gamma; + const PluginField* fields = fc->fields; + for (int i = 0; i < fc->nbFields; ++i) { + const char* attrName = fields[i].name; + if (!strcmp(attrName, "gamma")) { + assert(fields[i].type == PluginFieldType::kFLOAT32); + int size = fields[i].length; + gamma.reserve(size); + const auto* w = static_cast(fields[i].data); + for (int j = 0; j < size; j++) + { + gamma.push_back(*w); + w++; + } + } + } + + PReluPlugin* obj = new PReluPlugin(gamma); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + + IPluginV2IOExt* PReluPluginCreator::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() + PReluPlugin* obj = new PReluPlugin(serialData, serialLength); + obj->setPluginNamespace(mNamespace.c_str()); + return obj; + } + +} + diff --git a/arcface/prelu.h b/arcface/prelu.h new file mode 100644 index 0000000..2a6c033 --- /dev/null +++ b/arcface/prelu.h @@ -0,0 +1,107 @@ +#ifndef _PRELU_PLUGIN_H +#define _PRELU_PLUGIN_H + +#include +#include +#include "NvInfer.h" + +namespace nvinfer1 +{ + class PReluPlugin: public IPluginV2IOExt + { + public: + PReluPlugin(const std::vector& gamma); + PReluPlugin(const void* data, size_t length); + + ~PReluPlugin(); + + 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; + std::vector gamma_; + const char* mPluginNamespace; + }; + + class PReluPluginCreator : public IPluginCreator + { + public: + PReluPluginCreator(); + + ~PReluPluginCreator() 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; + }; +}; +#endif