add crnn

README.md

@@ -10,6 +10,7 @@ All the models are implemented in pytorch or mxnet first, and export a weights f
 
 ## News
 
+- `13 Sep 2020`. Add crnn, and got 1000fps on GTX1080.
 - `7 Sep 2020`. Implement retinaface(mobilenet0.25), and got 333fps on GTX1080.
 - `28 Aug 2020`. [BaofengZan](https://github.com/BaofengZan) added a tutorial for compiling and running tensorrtx on windows.
 - `16 Aug 2020`. [upczww](https://github.com/upczww) added a python wrapper for yolov5.
@@ -64,6 +65,7 @@ Following models are implemented.
 |[arcface](./arcface)| LResNet50E-IR, weights from [deepinsight/insightface](https://github.com/deepinsight/insightface) |
 |[retinafaceAntiCov](./retinafaceAntiCov)| mobilenet0.25, weights from [deepinsight/insightface](https://github.com/deepinsight/insightface), retinaface anti-COVID-19, detect face and mask attribute |
 |[dbnet](./dbnet)| Scene Text Detection, weights from [BaofengZan/DBNet.pytorch](https://github.com/BaofengZan/DBNet.pytorch) |
+|[crnn](./crnn)| pytorch implementation from [meijieru/crnn.pytorch](https://github.com/meijieru/crnn.pytorch) |
 
 ## Tricky Operations
 
@@ -87,6 +89,7 @@ Some tricky operations encountered in these models, already solved, but might ha
 |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 |
 |HardSwish| hard_swish = x * hard_sigmoid, used in yolov5 v3.0 |
+|LSTM| Implemented pytorch nn.LSTM() with tensorrt api |
 
 ## Speed Benchmark
 
@@ -107,6 +110,7 @@ Some tricky operations encountered in these models, already solved, but might ha
 | RetinaFace(resnet50) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 90 |
 | RetinaFace(mobilenet0.25) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 333 |
 | ArcFace(LResNet50E-IR) | Xeon E5-2620/GTX1080 | 1 | FP32 | 112x112 | 333 |
+| CRNN | Xeon E5-2620/GTX1080 | 1 | FP32 | 32x100 | 1000 |
 
 Help wanted, if you got speed results, please add an issue or PR.
 

crnn/CMakeLists.txt

@@ -0,0 +1,33 @@
+cmake_minimum_required(VERSION 2.6)
+
+project(crnn)
+
+add_definitions(-std=c++11)
+
+option(CUDA_USE_STATIC_CUDA_RUNTIME OFF)
+set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_BUILD_TYPE Debug)
+
+find_package(CUDA REQUIRED)
+
+include_directories(${PROJECT_SOURCE_DIR}/include)
+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()
+
+find_package(OpenCV)
+include_directories(OpenCV_INCLUDE_DIRS)
+
+add_executable(crnn ${PROJECT_SOURCE_DIR}/crnn.cpp)
+target_link_libraries(crnn nvinfer)
+target_link_libraries(crnn cudart)
+target_link_libraries(crnn ${OpenCV_LIBS})
+
+add_definitions(-O2 -pthread)
+

crnn/README.md

@@ -0,0 +1,44 @@
+# crnn
+
+The Pytorch implementation is [meijieru/crnn.pytorch](https://github.com/meijieru/crnn.pytorch).
+
+## How to Run
+
+```
+1. generate crnn.wts from pytorch
+
+git clone https://github.com/wang-xinyu/tensorrtx.git
+git clone https://github.com/meijieru/crnn.pytorch.git
+// download its weights 'crnn.pth'
+// copy tensorrtx/crnn/genwts.py into crnn.pytorch/
+// go to crnn.pytorch/
+python genwts.py
+// a file 'crnn.wts' will be generated.
+
+2. build tensorrtx/crnn and run
+
+// put crnn.wts into tensorrtx/crnn
+// go to tensorrtx/crnn
+mkdir build
+cd build
+cmake ..
+make
+sudo ./crnn -s  // serialize model to plan file i.e. 'crnn.engine'
+// copy crnn.pytorch/data/demo.png here
+sudo ./crnn -d  // deserialize plan file and run inference
+
+3. check the output as follows:
+
+raw: a-----v--a-i-l-a-bb-l-e---
+sim: available
+
+```
+
+## More Information
+
+See the readme in [home page.](https://github.com/wang-xinyu/tensorrtx)
+
+## Acknowledgment
+
+Thanks for the donation for this crnn tensorrt implementation from @雍.
+

crnn/crnn.cpp

@@ -0,0 +1,418 @@
+#include <iostream>
+#include <chrono>
+#include <map>
+#include <opencv2/opencv.hpp>
+#include "NvInfer.h"
+#include "cuda_runtime_api.h"
+#include "logging.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
+
+// stuff we know about the network and the input/output blobs
+static const int INPUT_H = 32;
+static const int INPUT_W = 100;
+static const int OUTPUT_SIZE = 26 * 37;
+const char* INPUT_BLOB_NAME = "data";
+const char* OUTPUT_BLOB_NAME = "prob";
+static Logger gLogger;
+
+const int ks[] = {3, 3, 3, 3, 3, 3, 2};
+const int ps[] = {1, 1, 1, 1, 1, 1, 0};
+const int ss[] = {1, 1, 1, 1, 1, 1, 1};
+const int nm[] = {64, 128, 256, 256, 512, 512, 512};
+const std::string alphabet = "-0123456789abcdefghijklmnopqrstuvwxyz";
+
+using namespace nvinfer1;
+
+std::string strDecode(std::vector<int>& preds, bool raw) {
+    std::string str;
+    if (raw) {
+        for (auto v: preds) {
+            str.push_back(alphabet[v]);
+        }
+    } else {
+        for (size_t i = 0; i < preds.size(); i++) {
+            if (preds[i] == 0 || (i > 0 && preds[i - 1] == preds[i])) continue;
+            str.push_back(alphabet[preds[i]]);
+        }
+    }
+    return str;
+}
+
+// TensorRT weight files have a simple space delimited format:
+// [type] [size] <data x size in hex>
+std::map<std::string, Weights> loadWeights(const std::string file) {
+    std::cout << "Loading weights: " << file << std::endl;
+    std::map<std::string, Weights> weightMap;
+
+    // Open weights file
+    std::ifstream input(file);
+    assert(input.is_open() && "Unable to load weight file. please check if the .wts file path is right!!!!!!");
+
+    // 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<uint32_t*>(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<std::string, Weights>& weightMap, ITensor& input, std::string lname, float eps) {
+    float *gamma = (float*)weightMap[lname + ".weight"].values;
+    float *beta = (float*)weightMap[lname + ".bias"].values;
+    float *mean = (float*)weightMap[lname + ".running_mean"].values;
+    float *var = (float*)weightMap[lname + ".running_var"].values;
+    int len = weightMap[lname + ".running_var"].count;
+
+    float *scval = reinterpret_cast<float*>(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<float*>(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<float*>(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* convRelu(INetworkDefinition *network, std::map<std::string, Weights>& weightMap, ITensor& input, int i, bool use_bn = false) {
+    int nOut = nm[i];
+    IConvolutionLayer* conv = network->addConvolutionNd(input, nOut, DimsHW{ks[i], ks[i]}, weightMap["cnn.conv" + std::to_string(i) + ".weight"], weightMap["cnn.conv" + std::to_string(i) + ".bias"]);
+    assert(conv);
+    conv->setStrideNd(DimsHW{ss[i], ss[i]});
+    conv->setPaddingNd(DimsHW{ps[i], ps[i]});
+    ILayer *tmp = conv;
+    if (use_bn) {
+        tmp = addBatchNorm2d(network, weightMap, *conv->getOutput(0), "cnn.batchnorm" + std::to_string(i), 1e-5);
+    }
+    auto relu = network->addActivation(*tmp->getOutput(0), ActivationType::kRELU);
+    assert(relu);
+    return relu;
+}
+
+void splitLstmWeights(std::map<std::string, Weights>& weightMap, std::string lname) {
+    int weight_size = weightMap[lname].count;
+    for (int i = 0; i < 4; i++) {
+        Weights wt{DataType::kFLOAT, nullptr, 0};
+        wt.count = weight_size / 4;
+        float *val = reinterpret_cast<float*>(malloc(sizeof(float) * wt.count));
+        memcpy(val, (float*)weightMap[lname].values + wt.count * i, sizeof(float) * wt.count);
+        wt.values = val;
+        weightMap[lname + std::to_string(i)] = wt;
+    }
+}
+
+ILayer* addLSTM(INetworkDefinition *network, std::map<std::string, Weights>& weightMap, ITensor& input, int nHidden, std::string lname) {
+    splitLstmWeights(weightMap, lname + ".weight_ih_l0");
+    splitLstmWeights(weightMap, lname + ".weight_hh_l0");
+    splitLstmWeights(weightMap, lname + ".bias_ih_l0");
+    splitLstmWeights(weightMap, lname + ".bias_hh_l0");
+    splitLstmWeights(weightMap, lname + ".weight_ih_l0_reverse");
+    splitLstmWeights(weightMap, lname + ".weight_hh_l0_reverse");
+    splitLstmWeights(weightMap, lname + ".bias_ih_l0_reverse");
+    splitLstmWeights(weightMap, lname + ".bias_hh_l0_reverse");
+    Dims dims = input.getDimensions();
+    std::cout << "lstm input shape: " << dims.nbDims << " [" << dims.d[0] << " " << dims.d[1] << " " << dims.d[2] << "]"<< std::endl;
+    auto lstm = network->addRNNv2(input, 1, nHidden, dims.d[1], RNNOperation::kLSTM);
+    lstm->setDirection(RNNDirection::kBIDIRECTION);
+    lstm->setWeightsForGate(0, RNNGateType::kINPUT, true, weightMap[lname + ".weight_ih_l00"]);
+    lstm->setWeightsForGate(0, RNNGateType::kFORGET, true, weightMap[lname + ".weight_ih_l01"]);
+    lstm->setWeightsForGate(0, RNNGateType::kCELL, true, weightMap[lname + ".weight_ih_l02"]);
+    lstm->setWeightsForGate(0, RNNGateType::kOUTPUT, true, weightMap[lname + ".weight_ih_l03"]);
+
+    lstm->setWeightsForGate(0, RNNGateType::kINPUT, false, weightMap[lname + ".weight_hh_l00"]);
+    lstm->setWeightsForGate(0, RNNGateType::kFORGET, false, weightMap[lname + ".weight_hh_l01"]);
+    lstm->setWeightsForGate(0, RNNGateType::kCELL, false, weightMap[lname + ".weight_hh_l02"]);
+    lstm->setWeightsForGate(0, RNNGateType::kOUTPUT, false, weightMap[lname + ".weight_hh_l03"]);
+
+    lstm->setBiasForGate(0, RNNGateType::kINPUT, true, weightMap[lname + ".bias_ih_l00"]);
+    lstm->setBiasForGate(0, RNNGateType::kFORGET, true, weightMap[lname + ".bias_ih_l01"]);
+    lstm->setBiasForGate(0, RNNGateType::kCELL, true, weightMap[lname + ".bias_ih_l02"]);
+    lstm->setBiasForGate(0, RNNGateType::kOUTPUT, true, weightMap[lname + ".bias_ih_l03"]);
+
+    lstm->setBiasForGate(0, RNNGateType::kINPUT, false, weightMap[lname + ".bias_hh_l00"]);
+    lstm->setBiasForGate(0, RNNGateType::kFORGET, false, weightMap[lname + ".bias_hh_l01"]);
+    lstm->setBiasForGate(0, RNNGateType::kCELL, false, weightMap[lname + ".bias_hh_l02"]);
+    lstm->setBiasForGate(0, RNNGateType::kOUTPUT, false, weightMap[lname + ".bias_hh_l03"]);
+
+    lstm->setWeightsForGate(1, RNNGateType::kINPUT, true, weightMap[lname + ".weight_ih_l0_reverse0"]);
+    lstm->setWeightsForGate(1, RNNGateType::kFORGET, true, weightMap[lname + ".weight_ih_l0_reverse1"]);
+    lstm->setWeightsForGate(1, RNNGateType::kCELL, true, weightMap[lname + ".weight_ih_l0_reverse2"]);
+    lstm->setWeightsForGate(1, RNNGateType::kOUTPUT, true, weightMap[lname + ".weight_ih_l0_reverse3"]);
+
+    lstm->setWeightsForGate(1, RNNGateType::kINPUT, false, weightMap[lname + ".weight_hh_l0_reverse0"]);
+    lstm->setWeightsForGate(1, RNNGateType::kFORGET, false, weightMap[lname + ".weight_hh_l0_reverse1"]);
+    lstm->setWeightsForGate(1, RNNGateType::kCELL, false, weightMap[lname + ".weight_hh_l0_reverse2"]);
+    lstm->setWeightsForGate(1, RNNGateType::kOUTPUT, false, weightMap[lname + ".weight_hh_l0_reverse3"]);
+
+    lstm->setBiasForGate(1, RNNGateType::kINPUT, true, weightMap[lname + ".bias_ih_l0_reverse0"]);
+    lstm->setBiasForGate(1, RNNGateType::kFORGET, true, weightMap[lname + ".bias_ih_l0_reverse1"]);
+    lstm->setBiasForGate(1, RNNGateType::kCELL, true, weightMap[lname + ".bias_ih_l0_reverse2"]);
+    lstm->setBiasForGate(1, RNNGateType::kOUTPUT, true, weightMap[lname + ".bias_ih_l0_reverse3"]);
+
+    lstm->setBiasForGate(1, RNNGateType::kINPUT, false, weightMap[lname + ".bias_hh_l0_reverse0"]);
+    lstm->setBiasForGate(1, RNNGateType::kFORGET, false, weightMap[lname + ".bias_hh_l0_reverse1"]);
+    lstm->setBiasForGate(1, RNNGateType::kCELL, false, weightMap[lname + ".bias_hh_l0_reverse2"]);
+    lstm->setBiasForGate(1, RNNGateType::kOUTPUT, false, weightMap[lname + ".bias_hh_l0_reverse3"]);
+    return lstm;
+}
+
+// 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 {C, INPUT_H, INPUT_W} with name INPUT_BLOB_NAME
+    ITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims3{1, INPUT_H, INPUT_W});
+    assert(data);
+
+    std::map<std::string, Weights> weightMap = loadWeights("../crnn.wts");
+
+    // cnn
+    auto x = convRelu(network, weightMap, *data, 0);
+    auto p = network->addPoolingNd(*x->getOutput(0), PoolingType::kMAX, DimsHW{2, 2});
+    p->setStrideNd(DimsHW{2, 2});
+    x = convRelu(network, weightMap, *p->getOutput(0), 1);
+    p = network->addPoolingNd(*x->getOutput(0), PoolingType::kMAX, DimsHW{2, 2});
+    p->setStrideNd(DimsHW{2, 2});
+    x = convRelu(network, weightMap, *p->getOutput(0), 2, true);
+    x = convRelu(network, weightMap, *x->getOutput(0), 3);
+    p = network->addPoolingNd(*x->getOutput(0), PoolingType::kMAX, DimsHW{2, 2});
+    p->setStrideNd(DimsHW{2, 1});
+    p->setPaddingNd(DimsHW{0, 1});
+    x = convRelu(network, weightMap, *p->getOutput(0), 4, true);
+    x = convRelu(network, weightMap, *x->getOutput(0), 5);
+    p = network->addPoolingNd(*x->getOutput(0), PoolingType::kMAX, DimsHW{2, 2});
+    p->setStrideNd(DimsHW{2, 1});
+    p->setPaddingNd(DimsHW{0, 1});
+    x = convRelu(network, weightMap, *p->getOutput(0), 6, true);
+
+    auto sfl = network->addShuffle(*x->getOutput(0));
+    sfl->setFirstTranspose(Permutation{1, 2, 0});
+
+    // rnn
+    auto lstm0 = addLSTM(network, weightMap, *sfl->getOutput(0), 256, "rnn.0.rnn");
+    auto sfl0 = network->addShuffle(*lstm0->getOutput(0));
+    sfl0->setReshapeDimensions(Dims4{26, 1, 1, 512});
+    auto fc0 = network->addFullyConnected(*sfl0->getOutput(0), 256, weightMap["rnn.0.embedding.weight"], weightMap["rnn.0.embedding.bias"]);
+
+    sfl = network->addShuffle(*fc0->getOutput(0));
+    sfl->setFirstTranspose(Permutation{2, 3, 0, 1});
+    sfl->setReshapeDimensions(Dims3{1, 26, 256});
+
+    auto lstm1 = addLSTM(network, weightMap, *sfl->getOutput(0), 256, "rnn.1.rnn");
+    auto sfl1 = network->addShuffle(*lstm1->getOutput(0));
+    sfl1->setReshapeDimensions(Dims4{26, 1, 1, 512});
+    auto fc1 = network->addFullyConnected(*sfl1->getOutput(0), 37, weightMap["rnn.1.embedding.weight"], weightMap["rnn.1.embedding.bias"]);
+    Dims dims = fc1->getOutput(0)->getDimensions();
+    std::cout << "fc1 shape " << dims.d[0] << " " << dims.d[1] << " " << dims.d[2] << std::endl;
+
+    fc1->getOutput(0)->setName(OUTPUT_BLOB_NAME);
+    network->markOutput(*fc1->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, cudaStream_t& stream, void **buffers, float* input, float* output, int batchSize) {
+    // DMA input batch data to device, infer on the batch asynchronously, and DMA output back to host
+    CHECK(cudaMemcpyAsync(buffers[0], input, batchSize * 1 * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream));
+    context.enqueue(batchSize, buffers, stream, nullptr);
+    CHECK(cudaMemcpyAsync(output, buffers[1], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));
+    cudaStreamSynchronize(stream);
+}
+
+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("crnn.engine", std::ios::binary);
+        if (!p) {
+            std::cerr << "could not open plan output file" << std::endl;
+            return -1;
+        }
+        p.write(reinterpret_cast<const char*>(modelStream->data()), modelStream->size());
+        modelStream->destroy();
+        return 0;
+    } else if (argc == 2 && std::string(argv[1]) == "-d") {
+        std::ifstream file("crnn.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 << "./crnn -s  // serialize model to plan file" << std::endl;
+        std::cerr << "./crnn -d ../samples  // deserialize plan file and run inference" << std::endl;
+        return -1;
+    }
+
+    // prepare input data ---------------------------
+    static float data[BATCH_SIZE * 1 * INPUT_H * INPUT_W];
+    //for (int i = 0; i < 1 * 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;
+    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);
+    assert(inputIndex == 0);
+    assert(outputIndex == 1);
+    // Create GPU buffers on device
+    CHECK(cudaMalloc(&buffers[inputIndex], BATCH_SIZE * 1 * INPUT_H * INPUT_W * sizeof(float)));
+    CHECK(cudaMalloc(&buffers[outputIndex], BATCH_SIZE * OUTPUT_SIZE * sizeof(float)));
+    // Create stream
+    cudaStream_t stream;
+    CHECK(cudaStreamCreate(&stream));
+
+    cv::Mat img = cv::imread("demo.png");
+    if (img.empty()) {
+        std::cerr << "demo.png not found !!!" << std::endl;
+        return -1;
+    }
+    cv::cvtColor(img, img, CV_BGR2GRAY);
+    cv::resize(img, img, cv::Size(INPUT_W, INPUT_H));
+    for (int i = 0; i < INPUT_H * INPUT_W; i++) {
+        data[i] = ((float)img.at<uchar>(i) / 255.0 - 0.5) * 2.0;
+    }
+
+    // Run inference
+    auto start = std::chrono::system_clock::now();
+    doInference(*context, stream, buffers, data, prob, BATCH_SIZE);
+    auto end = std::chrono::system_clock::now();
+    std::cout << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;
+
+    std::vector<int> preds;
+    for (int i = 0; i < 26; i++) {
+        int maxj = 0;
+        for (int j = 1; j < 37; j++) {
+            if (prob[37 * i + j] > prob[37 * i + maxj]) maxj = j;
+        }
+        preds.push_back(maxj);
+    }
+    std::cout << "raw: " << strDecode(preds, true) << std::endl;
+    std::cout << "sim: " << strDecode(preds, false) << std::endl;
+
+    // Release stream and buffers
+    cudaStreamDestroy(stream);
+    CHECK(cudaFree(buffers[inputIndex]));
+    CHECK(cudaFree(buffers[outputIndex]));
+    // Destroy the engine
+    context->destroy();
+    engine->destroy();
+    runtime->destroy();
+
+    // Print histogram of the output distribution
+    //std::cout << "\nOutput:\n\n";
+    //for (unsigned int i = 0; i < OUTPUT_SIZE; i++)
+    //{
+    //    std::cout << prob[i] << ", ";
+    //    if (i % 10 == 0) std::cout << std::endl;
+    //}
+    //std::cout << std::endl;
+
+    return 0;
+}

crnn/genwts.py

@@ -0,0 +1,35 @@
+import torch
+from torch.autograd import Variable
+import utils
+import models.crnn as crnn
+import struct
+
+model_path = './data/crnn.pth'
+
+model = crnn.CRNN(32, 1, 37, 256)
+if torch.cuda.is_available():
+    model = model.cuda()
+print('loading pretrained model from %s' % model_path)
+model.load_state_dict(torch.load(model_path))
+
+image = torch.ones(1, 1, 32, 100)
+if torch.cuda.is_available():
+    image = image.cuda()
+
+model.eval()
+print(model)
+print('image shape ', image.shape)
+preds = model(image)
+
+f = open("crnn.wts", 'w')
+f.write("{}\n".format(len(model.state_dict().keys())))
+for k,v in model.state_dict().items():
+    print('key: ', k)
+    print('value: ', v.shape)
+    vr = v.reshape(-1).cpu().numpy()
+    f.write("{} {}".format(k, len(vr)))
+    for vv in vr:
+        f.write(" ")
+        f.write(struct.pack(">f", float(vv)).hex())
+    f.write("\n")
+

crnn/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 <cassert>
+#include <ctime>
+#include <iomanip>
+#include <iostream>
+#include <ostream>
+#include <sstream>
+#include <string>
+
+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(&timestamp);
+            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