rcnn.cpp

#include <iostream>
#include <opencv2/opencv.hpp>
#include "backbone.hpp"
#include "RpnDecodePlugin.h"
#include "RpnNmsPlugin.h"
#include "RoiAlignPlugin.h"
#include "PredictorDecodePlugin.h"
#include "BatchedNmsPlugin.h"
#include "MaskRcnnInferencePlugin.h"
#include "calibrator.hpp"

#define DEVICE 0
#define BATCH_SIZE 1
#define BACKBONE_RESNETTYPE R50
// data
static const std::vector<float> PIXEL_MEAN = { 103.53, 116.28, 123.675 };
static const std::vector<float> PIXEL_STD = {1.0, 1.0, 1.0};
static constexpr float MIN_SIZE = 800.0;
static constexpr float MAX_SIZE = 1333.0;
static constexpr int NUM_CLASSES = 80;
static int INPUT_H;  // size of model input
static int INPUT_W;
static constexpr int INPUT_H_ = 480;  // size of original image, you can change it to arbitrary size
static constexpr int INPUT_W_ = 640;
static int X_LEFT_PAD;  // pad in preprocessImg
static int X_RIGHT_PAD;
static int Y_TOP_PAD;
static int Y_BOTTOM_PAD;
static int h_ori;  // used when h_ori is not equal to INPUT_H_
static int w_ori;
// backbone
static const int RES2_OUT_CHANNELS = (BACKBONE_RESNETTYPE == R18 ||
BACKBONE_RESNETTYPE == R34) ? 64 : 256;
// rpn
static const std::vector<float> ANCHOR_SIZES = { 32, 64, 128, 256, 512 };
static const std::vector<float> ASPECT_RATIOS = { 0.5, 1.0, 2.0 };
static constexpr int PRE_NMS_TOP_K_TEST = 6000;
static constexpr float RPN_NMS_THRESH = 0.7;
static constexpr int POST_NMS_TOPK = 1000;
// roialign
static constexpr int STRIDES = 16;
static constexpr int SAMPLING_RATIO = 0;
static constexpr int POOLER_RESOLUTION = 14;
// roihead
static constexpr float NMS_THRESH_TEST = 0.5;
static constexpr int DETECTIONS_PER_IMAGE = 100;
static constexpr float SCORE_THRESH = 0.6;
static const std::vector<float> BBOX_REG_WEIGHTS = { 10.0, 10.0, 5.0, 5.0 };
static bool MASK_ON = false;

static const char* INPUT_NODE_NAME = "images";
static const std::vector<std::string> OUTPUT_NAMES = { "scores", "boxes",
"labels", "masks" };

std::vector<float> GenerateAnchors(const std::vector<float>& anchor_sizes,
const std::vector<float>& aspect_ratios) {
    std::vector<float> res;
    for (auto as : anchor_sizes) {
        float area = as * as;
        for (auto ar : aspect_ratios) {
            float w = sqrt(area / ar);
            float h = ar * w;
            res.push_back(-w / 2.0);
            res.push_back(-h / 2.0);
            res.push_back(w / 2.0);
            res.push_back(h / 2.0);
        }
    }
    return res;
}

// transpose && resize && normalization && padding
ITensor* DataPreprocess(INetworkDefinition *network, ITensor& input) {

    // HWC->CHW
    auto channel_permute = network->addShuffle(input);
    assert(channel_permute);
    channel_permute->setFirstTranspose(Permutation{ 2, 0, 1 });

    // sub pixel mean
    auto pixel_mean = network->addConstant(Dims3{ 3, 1, 1 },
    Weights{ DataType::kFLOAT, PIXEL_MEAN.data(), 3 });
    assert(pixel_mean);
    auto sub = network->addElementWise(*channel_permute->getOutput(0),
    *pixel_mean->getOutput(0), ElementWiseOperation::kSUB);
    assert(sub);
    auto pixel_std = network->addConstant(Dims3{ 3, 1, 1 }, Weights{DataType::kFLOAT, PIXEL_STD.data(), 3});
    assert(pixel_std);
    auto div = network->addElementWise(*sub->getOutput(0), *pixel_std->getOutput(0), ElementWiseOperation::kDIV);
    assert(div);

    return div->getOutput(0);
}

ITensor* RPN(INetworkDefinition *network,
std::map<std::string, Weights>& weightMap, ITensor& features) {
    int num_anchors = ANCHOR_SIZES.size() * ASPECT_RATIOS.size();
    int box_dim = 4;

    // rpn head conv
    auto rpn_head_conv = network->addConvolutionNd(features, features.getDimensions().d[0], DimsHW{ 3, 3 },
    weightMap["proposal_generator.rpn_head.conv.weight"],
    weightMap["proposal_generator.rpn_head.conv.bias"]);
    assert(rpn_head_conv);
    rpn_head_conv->setStrideNd(DimsHW{ 1, 1 });
    rpn_head_conv->setPaddingNd(DimsHW{ 1, 1 });
    auto rpn_head_relu = network->addActivation(*rpn_head_conv->getOutput(0), ActivationType::kRELU);
    assert(rpn_head_relu);

    // objectness logits
    auto rpn_head_logits = network->addConvolutionNd(*rpn_head_relu->getOutput(0), num_anchors, DimsHW{ 1, 1 },
    weightMap["proposal_generator.rpn_head.objectness_logits.weight"],
    weightMap["proposal_generator.rpn_head.objectness_logits.bias"]);
    assert(rpn_head_logits);
    rpn_head_logits->setStrideNd(DimsHW{ 1, 1 });

    // anchor deltas
    auto rpn_head_deltas = network->addConvolutionNd(*rpn_head_relu->getOutput(0), num_anchors * box_dim,
    DimsHW{ 1, 1 },
    weightMap["proposal_generator.rpn_head.anchor_deltas.weight"],
    weightMap["proposal_generator.rpn_head.anchor_deltas.bias"]);
    assert(rpn_head_deltas);
    auto rpn_head_deltas_dim = rpn_head_deltas->getOutput(0)->getDimensions();
    rpn_head_deltas->setStrideNd(DimsHW{ 1, 1 });

    auto anchors = GenerateAnchors(ANCHOR_SIZES, ASPECT_RATIOS);
    auto rpnDecodePlugin = RpnDecodePlugin(PRE_NMS_TOP_K_TEST, anchors, STRIDES, INPUT_H, INPUT_W);
    std::vector<ITensor*> faster_decode_inputs = { rpn_head_logits->getOutput(0), rpn_head_deltas->getOutput(0) };
    auto rpnDecodeLayer = network->addPluginV2(faster_decode_inputs.data(), faster_decode_inputs.size(),
    rpnDecodePlugin);

    std::vector<ITensor*> nms_input = { rpnDecodeLayer->getOutput(0), rpnDecodeLayer->getOutput(1) };

    // nms
    auto nmsPlugin = RpnNmsPlugin(RPN_NMS_THRESH, POST_NMS_TOPK);
    auto nmsLayer = network->addPluginV2(nms_input.data(), nms_input.size(), nmsPlugin);
    return nmsLayer->getOutput(0);
}

ITensor* SharedRoiTransform(INetworkDefinition *network, std::map<std::string, Weights>& weightMap,
ITensor* proposals, ITensor* features, int num_proposals) {
    std::vector<ITensor*> roi_inputs = { proposals, features };
    auto roiAlignPlugin = RoiAlignPlugin(POOLER_RESOLUTION, 1 / static_cast<float>(STRIDES),
    SAMPLING_RATIO, num_proposals, features->getDimensions().d[0]);
    auto roiAlignLayer = network->addPluginV2(roi_inputs.data(), roi_inputs.size(), roiAlignPlugin);

    // res5
    /* same with https://github.com/facebookresearch/detectron2/
    blob/9246ebc3af1c023cfbdae77e5d976edbcf9a2933/detectron2/modeling/roi_heads/roi_heads.py#L430,
    use bottleneck here, so pass R50*/
    auto box_features = MakeStage(network, weightMap, "roi_heads.res5",
    *roiAlignLayer->getOutput(0), 3, R50,
    roiAlignLayer->getOutput(0)->getDimensions().d[1],
    512, RES2_OUT_CHANNELS * 8, 2);
    return box_features;
}

void BoxHead(INetworkDefinition *network, std::map<std::string, Weights>& weightMap, ITensor* proposals,
    ITensor* features, std::vector<ITensor*>& instances) {

    auto box_features = SharedRoiTransform(network, weightMap, proposals, features, POST_NMS_TOPK);
    auto box_features_mean = network->addReduce(*box_features, ReduceOperation::kAVG, 12, true);

    // score
    auto scores = network->addFullyConnected(*box_features_mean->getOutput(0), NUM_CLASSES + 1,
    weightMap["roi_heads.box_predictor.cls_score.weight"],
    weightMap["roi_heads.box_predictor.cls_score.bias"]);
    auto probs = network->addSoftMax(*scores->getOutput(0));

    auto probs_dim = probs->getOutput(0)->getDimensions();
    auto score_slice = network->addSlice(*probs->getOutput(0), Dims4{ 0, 0, 0, 0 },
    Dims4{ probs_dim.d[0], probs_dim.d[1] - 1, 1, 1 }, Dims4{ 1, 1, 1, 1 });

    auto proposal_deltas = network->addFullyConnected(*box_features_mean->getOutput(0), NUM_CLASSES * 4,
    weightMap["roi_heads.box_predictor.bbox_pred.weight"],
    weightMap["roi_heads.box_predictor.bbox_pred.bias"]);

    // decode
    std::vector<ITensor*> predictorDecodeInput = { score_slice->getOutput(0),
    proposal_deltas->getOutput(0), proposals };
    auto predictorDecodePlugin = PredictorDecodePlugin(probs_dim.d[0], INPUT_H, INPUT_W, BBOX_REG_WEIGHTS);
    auto predictorDecodeLayer = network->addPluginV2(predictorDecodeInput.data(),
    predictorDecodeInput.size(), predictorDecodePlugin);

    // nms
    std::vector<ITensor*> nmsInput = { predictorDecodeLayer->getOutput(0),
    predictorDecodeLayer->getOutput(1), predictorDecodeLayer->getOutput(2) };
    auto batchedNmsPlugin = BatchedNmsPlugin(NMS_THRESH_TEST, DETECTIONS_PER_IMAGE);
    auto batchedNmsLayer = network->addPluginV2(nmsInput.data(), nmsInput.size(), batchedNmsPlugin);

    // instances
    instances.push_back(batchedNmsLayer->getOutput(0));
    instances.push_back(batchedNmsLayer->getOutput(1));
    instances.push_back(batchedNmsLayer->getOutput(2));
}

void MaskHead(INetworkDefinition *network, std::map<std::string, Weights>& weightMap,
    ITensor* features, std::vector<ITensor*>& instances, int out_channels = 256) {

    auto mask_features = SharedRoiTransform(network, weightMap, instances[1], features, DETECTIONS_PER_IMAGE);

    // mask_fcn
    auto mask_deconv = network->addDeconvolutionNd(*mask_features, out_channels, DimsHW{ 2, 2 },
    weightMap["roi_heads.mask_head.deconv.weight"],
    weightMap["roi_heads.mask_head.deconv.bias"]);
    mask_deconv->setStrideNd(DimsHW{ 2, 2 });
    auto deconv_relu = network->addActivation(*mask_deconv->getOutput(0), ActivationType::kRELU);
    assert(deconv_relu);
    auto predictor = network->addConvolutionNd(*deconv_relu->getOutput(0), NUM_CLASSES, DimsHW{ 1, 1 },
    weightMap["roi_heads.mask_head.predictor.weight"],
    weightMap["roi_heads.mask_head.predictor.bias"]);
    predictor->setStrideNd(DimsHW{ 1, 1 });

    ITensor* masks;
    if (NUM_CLASSES == 1) {
        auto mask_probs_pred = network->addActivation(*predictor->getOutput(0), ActivationType::kSIGMOID);
        masks = mask_probs_pred->getOutput(0);
    } else {
        std::vector<ITensor*> mask_rcnn_inference_inputs = { instances[2], predictor->getOutput(0) };
        auto maskRcnnInferencePlugin = MaskRcnnInferencePlugin(DETECTIONS_PER_IMAGE, POOLER_RESOLUTION);
        auto maskRcnnInferenceLayer = network->addPluginV2(mask_rcnn_inference_inputs.data(),
        mask_rcnn_inference_inputs.size(), maskRcnnInferencePlugin);
        masks = maskRcnnInferenceLayer->getOutput(0);
    }
    instances.push_back(masks);
}

std::vector<ITensor*> ROIHeads(INetworkDefinition *network, std::map<std::string, Weights>& weightMap,
ITensor* proposals, ITensor* features) {
    std::vector<ITensor*> instances;

    // box head
    BoxHead(network, weightMap, proposals, features, instances);

    if (MASK_ON) {
        // mask head
        MaskHead(network, weightMap, features, instances);
    }

    return instances;
}

ICudaEngine* createEngine_rcnn(unsigned int maxBatchSize,
    const std::string& wtsfile, IBuilder* builder, IBuilderConfig* config, DataType dt,
    const std::string& quantizationType) {
    /*
    description: after fuse bn
    */
    INetworkDefinition* network = builder->createNetworkV2(0U);

    // Create input tensor of shape {INPUT_H, INPUT_W, 3} with name INPUT_BLOB_NAME
    ITensor* data = network->addInput(INPUT_NODE_NAME, dt, Dims3{ INPUT_H, INPUT_W, 3 });
    assert(data);

    // preprocess
    data = DataPreprocess(network, *data);
    std::map<std::string, Weights> weightMap;
    loadWeights(wtsfile, weightMap);

    // backbone
    ITensor* features = BuildResNet(network, weightMap, *data, BACKBONE_RESNETTYPE, 64, 64, RES2_OUT_CHANNELS);

    auto proposals = RPN(network, weightMap, *features);
    auto results = ROIHeads(network, weightMap, proposals, features);

    // build output
    for (int i = 0; i < results.size(); i++) {
        network->markOutput(*results[i]);
        results[i]->setName(OUTPUT_NAMES[i].c_str());
    }

    // build engine
    builder->setMaxBatchSize(maxBatchSize);
    config->setMaxWorkspaceSize(1ULL << 30);
    if (quantizationType == "fp32") {
    } else if (quantizationType == "fp16") {
        config->setFlag(BuilderFlag::kFP16);
    } else if (quantizationType == "int8") {
        std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
        assert(builder->platformHasFastInt8());
        config->setFlag(BuilderFlag::kINT8);
        Int8EntropyCalibrator2* calibrator = new Int8EntropyCalibrator2(1, INPUT_W, INPUT_H, "./coco_calib/",
        "int8calib.table", INPUT_NODE_NAME);
        config->setInt8Calibrator(calibrator);
    } else {
        throw("does not support model type");
    }

    std::cout << "Building engine, please wait for a while..." << std::endl;
    ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);
    std::cout << "Build engine successfully!" << std::endl;

    // destroy network
    network->destroy();

    // Release host memory
    for (auto& mem : weightMap) {
        delete[] mem.second.values;
    }
    return engine;
}

void BuildRcnnModel(unsigned int maxBatchSize, IHostMemory** modelStream, const std::string& wtsfile,
const std::string& quantizationType = "fp32") {
    // Create builder
    IBuilder* builder = createInferBuilder(gLogger);
    IBuilderConfig* config = builder->createBuilderConfig();

    ICudaEngine* engine = createEngine_rcnn(maxBatchSize,
        wtsfile, builder, config, DataType::kFLOAT, quantizationType);
    assert(engine != nullptr);

    // Serialize the engine
    (*modelStream) = engine->serialize();

    // Close everything down
    engine->destroy();
    builder->destroy();
}

void doInference(IExecutionContext& context, cudaStream_t& stream, std::vector<void*>& buffers,
std::vector<float>& input, std::vector<float*>& output) {
    CUDA_CHECK(cudaMemcpyAsync(buffers[0], input.data(), BATCH_SIZE * INPUT_H * INPUT_W * 3 * sizeof(float),
    cudaMemcpyHostToDevice, stream));

    context.enqueue(BATCH_SIZE, buffers.data(), stream, nullptr);

    CUDA_CHECK(cudaMemcpyAsync(output[0], buffers[1], BATCH_SIZE * DETECTIONS_PER_IMAGE * sizeof(float),
    cudaMemcpyDeviceToHost, stream));
    CUDA_CHECK(cudaMemcpyAsync(output[1], buffers[2], BATCH_SIZE * DETECTIONS_PER_IMAGE * 4 * sizeof(float),
    cudaMemcpyDeviceToHost, stream));
    CUDA_CHECK(cudaMemcpyAsync(output[2], buffers[3], BATCH_SIZE * DETECTIONS_PER_IMAGE * sizeof(float),
    cudaMemcpyDeviceToHost, stream));
    if (MASK_ON)
        CUDA_CHECK(cudaMemcpyAsync(output[3], buffers[4],
        BATCH_SIZE * DETECTIONS_PER_IMAGE * POOLER_RESOLUTION * POOLER_RESOLUTION * sizeof(float),
        cudaMemcpyDeviceToHost, stream));

    cudaStreamSynchronize(stream);
}

void calculateSize() {
    float ratio = MIN_SIZE / static_cast<float>(std::min(INPUT_H_, INPUT_W_));
    float newh = 0, neww = 0;
    if (INPUT_H_ < INPUT_W_) {
        newh = MIN_SIZE;
        neww = ratio * INPUT_W_;
    } else {
        newh = ratio * INPUT_H_;
        neww = MIN_SIZE;
    }
    if (std::max(newh, neww) > MAX_SIZE) {
        ratio = MAX_SIZE / static_cast<float>(std::max(newh, neww));
        newh = newh * ratio;
        neww = neww * ratio;
    }
    INPUT_H = static_cast<int>(newh + 0.5);
    INPUT_W = static_cast<int>(neww + 0.5);
}


bool parse_args(int argc, char** argv, std::string& wtsFile, std::string& engineFile, std::string& imgDir) {
    if (argc < 4) return false;
    if (std::string(argv[1]) == "-s") {
        wtsFile = std::string(argv[2]);
        engineFile = std::string(argv[3]);
    } else if (std::string(argv[1]) == "-d") {
        engineFile = std::string(argv[2]);
        imgDir = std::string(argv[3]);
    } else {
        return false;
    }
    if (argc >= 5 && std::string(argv[4]) == "m") MASK_ON = true;
    return true;
}

int main(int argc, char** argv) {

    // calculate size
    calculateSize();

    cudaSetDevice(DEVICE);

    std::string wtsFile = "";
    std::string engineFile = "";

    std::string imgDir;
    if (!parse_args(argc, argv, wtsFile, engineFile, imgDir)) {
        std::cerr << "arguments not right!" << std::endl;
        std::cerr << "./rcnn -s [.wts] [.engine] [m] // serialize model to plan file" << std::endl;
        std::cerr << "./rcnn -d [.engine] ../samples [m]  // deserialize plan file and run inference" << std::endl;
        return -1;
    }

    if (!wtsFile.empty()) {
        IHostMemory* modelStream{ nullptr };
        BuildRcnnModel(BATCH_SIZE, &modelStream, wtsFile, "fp32");
        assert(modelStream != nullptr);
        std::ofstream p(engineFile, 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;
    }

    // deserialize the .engine and run inference
    std::ifstream file(engineFile, std::ios::binary);
    if (!file.good()) {
        std::cerr << "read " << engineFile << " error!" << std::endl;
        return -1;
    }

    std::string trtModelStream;
    size_t modelSize{ 0 };
    file.seekg(0, file.end);
    modelSize = file.tellg();
    file.seekg(0, file.beg);
    trtModelStream.resize(modelSize);
    assert(!trtModelStream.empty());
    file.read(const_cast<char*>(trtModelStream.c_str()), modelSize);
    file.close();

    // build engine
    std::cout << "build engine" << std::endl;
    IRuntime* runtime = createInferRuntime(gLogger);
    assert(runtime != nullptr);
    ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream.c_str(), modelSize);
    assert(engine != nullptr);
    IExecutionContext* context = engine->createExecutionContext();
    assert(context != nullptr);
    runtime->destroy();

    cudaStream_t stream;
    CUDA_CHECK(cudaStreamCreate(&stream));

    // prepare input file
    std::vector<std::string> fileList;
    if (read_files_in_dir(imgDir.c_str(), fileList) < 0) {
        std::cerr << "read_files_in_dir failed." << std::endl;
        return -1;
    }

    // prepare input data
    std::vector<float> data(BATCH_SIZE * INPUT_H * INPUT_W * 3, 0);
    void *data_d, *scores_d, *boxes_d, *classes_d, *masks_d;
    CUDA_CHECK(cudaMalloc(&data_d, BATCH_SIZE * INPUT_H * INPUT_W * 3 * sizeof(float)));
    CUDA_CHECK(cudaMalloc(&scores_d, BATCH_SIZE * DETECTIONS_PER_IMAGE * sizeof(float)));
    CUDA_CHECK(cudaMalloc(&boxes_d, BATCH_SIZE * DETECTIONS_PER_IMAGE * 4 * sizeof(float)));
    CUDA_CHECK(cudaMalloc(&classes_d, BATCH_SIZE * DETECTIONS_PER_IMAGE * sizeof(float)));

    std::vector<float> scores_h(BATCH_SIZE * DETECTIONS_PER_IMAGE);
    std::vector<float> boxes_h(BATCH_SIZE * DETECTIONS_PER_IMAGE * 4);
    std::vector<float> classes_h(BATCH_SIZE * DETECTIONS_PER_IMAGE);
    std::vector<float> masks_h;

    std::vector<void*> buffers = { data_d, scores_d, boxes_d, classes_d };
    std::vector<float*> outputs = {scores_h.data(), boxes_h.data(), classes_h.data()};

    if (MASK_ON) {
        CUDA_CHECK(cudaMalloc(&masks_d,
        BATCH_SIZE * DETECTIONS_PER_IMAGE * POOLER_RESOLUTION * POOLER_RESOLUTION * sizeof(float)));
        masks_h.resize(BATCH_SIZE * DETECTIONS_PER_IMAGE * POOLER_RESOLUTION * POOLER_RESOLUTION);
        buffers.push_back(masks_d);
        outputs.push_back(masks_h.data());
    }

    int fcount = 0;
    int fileLen = fileList.size();
    for (int f = 0; f < fileLen; f++) {
        fcount++;
        if (fcount < BATCH_SIZE && f + 1 != fileLen) continue;

        for (int b = 0; b < fcount; b++) {
            cv::Mat img = cv::imread(imgDir + "/" + fileList[f - fcount + 1 + b]);
            h_ori = img.rows;
            w_ori = img.cols;
            img = preprocessImg(img, INPUT_W, INPUT_H, X_LEFT_PAD, X_RIGHT_PAD, Y_TOP_PAD, Y_BOTTOM_PAD);

            if (img.empty()) continue;
            for (int i = 0; i < INPUT_H * INPUT_W * 3; i++)
                data[b*INPUT_H * INPUT_W * 3 + i] = static_cast<float>(*(img.data + i));
        }

        // Run inference
        auto start = std::chrono::system_clock::now();

        doInference(*context, stream, buffers, data, outputs);

        auto end = std::chrono::system_clock::now();
        std::cout << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;

        float h_ratio = static_cast<float>(h_ori) / (INPUT_H - (Y_TOP_PAD + Y_BOTTOM_PAD));  // ratio of original image size to model input size
        float w_ratio = static_cast<float>(w_ori) / (INPUT_W - (X_LEFT_PAD + X_RIGHT_PAD));

        for (int b = 0; b < fcount; b++) {
            cv::Mat img = cv::imread(imgDir + "/" + fileList[f - fcount + 1 + b]);
            for (int i = 0; i < DETECTIONS_PER_IMAGE; i++) {
                if (scores_h[b * DETECTIONS_PER_IMAGE + i] > SCORE_THRESH) {
                    float x1 = (boxes_h[b * DETECTIONS_PER_IMAGE * 4 + i * 4 + 0] - X_LEFT_PAD) * w_ratio;
                    float y1 = (boxes_h[b * DETECTIONS_PER_IMAGE * 4 + i * 4 + 1] - Y_TOP_PAD) * h_ratio;
                    float x2 = (boxes_h[b * DETECTIONS_PER_IMAGE * 4 + i * 4 + 2] - X_LEFT_PAD) * w_ratio;
                    float y2 = (boxes_h[b * DETECTIONS_PER_IMAGE * 4 + i * 4 + 3] - Y_TOP_PAD) * h_ratio;
                    int label = classes_h[b * DETECTIONS_PER_IMAGE + i];
                    float score = scores_h[b * DETECTIONS_PER_IMAGE + i];
                    printf("boxes:[%.6f, %.6f, %.6f, %.6f] scores: %.4f label: %d \n", x1, y1, x2, y2, score, label);
                    cv::Rect r(x1, y1, x2 - x1, y2 - y1);
                    cv::rectangle(img, r, cv::Scalar(0x27, 0xC1, 0x36), 2);
                    cv::putText(img, std::to_string(label), cv::Point(r.x, r.y - 1), cv::FONT_HERSHEY_PLAIN, 1.2,
                    cv::Scalar(0xFF, 0xFF, 0xFF), 2);


                    if (MASK_ON) {
                        cv::Mat maskPart = cv::Mat::zeros(cv::Size(POOLER_RESOLUTION, POOLER_RESOLUTION), CV_32FC1);
                        memcpy(maskPart.data,
                          &masks_h[b * DETECTIONS_PER_IMAGE * POOLER_RESOLUTION * POOLER_RESOLUTION +
                          i * POOLER_RESOLUTION * POOLER_RESOLUTION],
                          POOLER_RESOLUTION * POOLER_RESOLUTION * sizeof(float));

                        cv::Rect r(cv::Point(floor(x1) - 1 < 0 ? 0 : floor(x1) - 1,
                                             floor(y1) - 1 < 0 ? 0 : floor(y1) - 1),
                                   cv::Point(ceil(x2) + 1 > w_ori ? w_ori : ceil(x2) + 1,
                                             ceil(y2) + 1 > h_ori ? h_ori : ceil(y2) + 1));
                        cv::resize(maskPart, maskPart, cv::Size(r.width, r.height));
                        cv::Mat curMask = cv::Mat::zeros(cv::Size(w_ori, h_ori), CV_8UC1);
                        cv::threshold(maskPart, maskPart, 0.5, 255, cv::THRESH_BINARY);
                        curMask(r) += maskPart;
                        std::vector<std::vector<cv::Point>> contours;
                        cv::findContours(curMask, contours, cv::RETR_EXTERNAL, cv::CHAIN_APPROX_NONE);
                        for (int c = 0; c < contours.size(); c++)
                            cv::drawContours(img, contours, c, cv::Scalar(0, 0, 255));
                    }
                }
            }
            cv::imwrite("_" + fileList[f - fcount + 1 + b], img);
        }
        fcount = 0;
    }

    cudaStreamDestroy(stream);
    CUDA_CHECK(cudaFree(data_d));
    CUDA_CHECK(cudaFree(scores_d));
    CUDA_CHECK(cudaFree(boxes_d));
    CUDA_CHECK(cudaFree(classes_d));
    if (MASK_ON) CUDA_CHECK(cudaFree(masks_d));
    context->destroy();
    engine->destroy();

    return 0;
}