yolov9 t s m (#1541)

* yolov9 t s m * change chinese
2024-06-21 15:44:24 +08:00 · 2024-06-21 15:44:24 +08:00 · e065c7508d
commit e065c7508d
parent cc725017a4
6 changed files with 830 additions and 229 deletions
--- a/yolov9/README.md
+++ b/yolov9/README.md
@ -7,22 +7,16 @@ The Pytorch implementation is [WongKinYiu/yolov9](https://github.com/WongKinYiu/
 <a href="https://github.com/WuxinrongY"><img src="https://avatars.githubusercontent.com/u/53141838?v=4?s=48" width="40px;" alt=""/></a>

 ## Progress
- [x] YOLOv9-c:
-  - [x] FP32
-  - [x] FP16
-  - [x] INT8
- [x] YOLOv9-e:
-  - [x] FP32
-  - [x] FP16
-  - [x] INT8
- [x] GELAN-c:
-  - [x] FP32
-  - [x] FP16
-  - [x] INT8
- [x] GELAN-e:
-  - [x] FP32
-  - [x] FP16
-  - [x] INT8
+- [x] YOLOv9-t
+- [x] YOLOv9-t-convert(gelan)
+- [x] YOLOv9-s
+- [x] YOLOv9-s-convert(gelan)
+- [x] YOLOv9-m
+- [x] YOLOv9-m-convert(gelan)
+- [x] YOLOv9-c
+- [x] YOLOv9-c-convert(gelan)
+- [x] YOLOv9-e
+- [x] YOLOv9-e-convert(gelan)

 ## Requirements

@ -35,8 +29,16 @@ The speed test is done on a desktop with R7-5700G CPU and RTX 4060Ti GPU. The in

 | frame  | Model | FP32 | FP16 | INT8 |
 | --- | --- | --- | --- | --- |
-| pytorch | YOLOv9-c | - | 15.5ms | - |
-| pytorch | YOLOv9-e | - | 19.7ms | - |
+| tensorrt | YOLOv5-n | -ms | 0.58ms | -ms |
+| tensorrt | YOLOv5-s | -ms | 0.90ms | -ms |
+| tensorrt | YOLOv5-m | -ms | 1.9ms | -ms |
+| tensorrt | YOLOv5-l | -ms | 2.8ms | -ms |
+| tensorrt | YOLOv5-x | -ms | 5.1ms | -ms |
+| tensorrt | YOLOv9-t-convert | -ms | 1.37ms | -ms |
+| tensorrt | YOLOv9-s | -ms | 1.78ms | -ms |
+| tensorrt | YOLOv9-s-convert | -ms | 1.78ms | -ms |
+| tensorrt | YOLOv9-m | -ms | 3.1ms | -ms |
+| tensorrt | YOLOv9-m-convert | -ms | 2.8ms | -ms |
 | tensorrt | YOLOv9-c | 13.5ms | 4.6ms | 3.0ms |
 | tensorrt | YOLOv9-e | 8.3ms | 3.2ms | 2.15ms |

--- a/yolov9/demo.cpp
+++ b/yolov9/demo.cpp
@ -19,17 +19,32 @@ void serialize_engine(unsigned int max_batchsize, std::string& wts_name, std::st

    // Create model to populate the network, then set the outputs and create an engine
    IHostMemory* serialized_engine = nullptr;
-    if (sub_type == "e") {
-        serialized_engine = build_engine_yolov9_e(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
+    if (sub_type == "t") {
+        serialized_engine = build_engine_yolov9_t(max_batchsize, builder, config, DataType::kFLOAT, wts_name, false);
+    } else if (sub_type == "s") {
+        serialized_engine = build_engine_yolov9_s(max_batchsize, builder, config, DataType::kFLOAT, wts_name, false);
+    } else if (sub_type == "m") {
+        serialized_engine = build_engine_yolov9_m(max_batchsize, builder, config, DataType::kFLOAT, wts_name, false);
    } else if (sub_type == "c") {
        serialized_engine = build_engine_yolov9_c(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
-    } else if (sub_type == "ge") {
-        serialized_engine = build_engine_gelan_e(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
+    } else if (sub_type == "e") {
+        serialized_engine = build_engine_yolov9_e(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
+    }
+
+    else if (sub_type == "gt") {
+        serialized_engine = build_engine_yolov9_t(max_batchsize, builder, config, DataType::kFLOAT, wts_name, true);
+    } else if (sub_type == "gs") {
+        serialized_engine = build_engine_yolov9_s(max_batchsize, builder, config, DataType::kFLOAT, wts_name, true);
+    } else if (sub_type == "gm") {
+        serialized_engine = build_engine_yolov9_m(max_batchsize, builder, config, DataType::kFLOAT, wts_name, true);
    } else if (sub_type == "gc") {
        serialized_engine = build_engine_gelan_c(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
+    } else if (sub_type == "ge") {
+        serialized_engine = build_engine_gelan_e(max_batchsize, builder, config, DataType::kFLOAT, wts_name);
    } else {
        return;
    }
+
    assert(serialized_engine != nullptr);

    std::ofstream p(engine_name, std::ios::binary);
@ -114,19 +129,19 @@ int main(int argc, char** argv) {
    cudaSetDevice(kGpuId);

    std::string wts_name = "";
-    std::string engine_name = "";
-    std::string img_dir = "";
-    std::string sub_type = "";
+    std::string engine_name = "../yolov9-m-converted.engine";
+    std::string img_dir = "../images";
+    std::string sub_type = "m";
    // speed test or inference
-    // const int speed_test_iter = 1000;
-    const int speed_test_iter = 1;
+    const int speed_test_iter = 1000;
+    // const int speed_test_iter = 1;

-    if (!parse_args(argc, argv, wts_name, engine_name, img_dir, sub_type)) {
-        std::cerr << "Arguments not right!" << std::endl;
-        std::cerr << "./yolov9 -s [.wts] [.engine] [c/e/gc/ge]  // serialize model to plan file" << std::endl;
-        std::cerr << "./yolov9 -d [.engine] ../samples  // deserialize plan file and run inference" << std::endl;
-        return -1;
-    }
+    // if (!parse_args(argc, argv, wts_name, engine_name, img_dir, sub_type)) {
+    //     std::cerr << "Arguments not right!" << std::endl;
+    //     std::cerr << "./yolov9 -s [.wts] [.engine] [s/m/c/e/gt/gs/gm/gc/ge]  // serialize model to plan file" << std::endl;
+    //     std::cerr << "./yolov9 -d [.engine] ../samples  // deserialize plan file and run inference" << std::endl;
+    //     return -1;
+    // }

    // Create a model using the API directly and serialize it to a file
    if (!wts_name.empty()) {
--- a/yolov9/include/block.h
+++ b/yolov9/include/block.h
@ -22,10 +22,14 @@ std::vector<std::vector<float>> getAnchors(std::map<std::string, Weights>& weigh
 // ----------------------------------------------------------------
 nvinfer1::ILayer* convBnSiLU(nvinfer1::INetworkDefinition* network, std::map<std::string, nvinfer1::Weights>& weightMap,
                             nvinfer1::ITensor& input, int ch, int k, int s, int p, std::string lname, int g = 1);
+ILayer* ELAN1(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c1, int c2,
+              int c3, int c4, std::string lname);
 ILayer* RepNCSPELAN4(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c1,
                     int c2, int c3, int c4, int c5, std::string lname);
 ILayer* ADown(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c2,
              std::string lname);
+ILayer* AConv(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c2,
+              std::string lname);
 std::vector<ILayer*> CBLinear(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input,
                              std::vector<int> c2s, int k, int s, int p, int g, std::string lname);
 ILayer* CBFuse(INetworkDefinition* network, std::vector<std::vector<ILayer*>> input, std::vector<int> idx,
--- a/yolov9/include/model.h
+++ b/yolov9/include/model.h
@ -2,16 +2,32 @@

 #include <NvInfer.h>
 #include <string>
-nvinfer1::IHostMemory* build_engine_yolov9_e(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+// yolov9
+nvinfer1::IHostMemory* build_engine_yolov9_t(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
                                             nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
-                                             std::string& wts_name);
+                                             std::string& wts_name, bool isConvert = false);
+nvinfer1::IHostMemory* build_engine_yolov9_s(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+                                             nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
+                                             std::string& wts_name, bool isConvert = false);
+nvinfer1::IHostMemory* build_engine_yolov9_m(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+                                             nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
+                                             std::string& wts_name, bool isConvert = false);
 nvinfer1::IHostMemory* build_engine_yolov9_c(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
                                             nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
                                             std::string& wts_name);
-
-nvinfer1::IHostMemory* build_engine_gelan_e(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+nvinfer1::IHostMemory* build_engine_yolov9_e(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+                                             nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
+                                             std::string& wts_name);
+// gelan
+nvinfer1::IHostMemory* build_engine_gelan_t(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
+                                            std::string& wts_name);
+nvinfer1::IHostMemory* build_engine_gelan_m(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
                                            std::string& wts_name);
 nvinfer1::IHostMemory* build_engine_gelan_c(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
                                            std::string& wts_name);
+nvinfer1::IHostMemory* build_engine_gelan_e(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
+                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
+                                            std::string& wts_name);
--- a/yolov9/src/block.cpp
+++ b/yolov9/src/block.cpp
@ -204,11 +204,34 @@ ILayer* RepNCSP(INetworkDefinition* network, std::map<std::string, Weights>& wei
    auto cv3 = convBnSiLU(network, weightMap, *cat->getOutput(0), c2, 1, 1, 0, lname + ".cv3", 1);
    return cv3;
 }
+
+ILayer* ELAN1(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c1, int c2,
+              int c3, int c4, std::string lname) {
+    auto cv1 = convBnSiLU(network, weightMap, input, c3, 1, 1, 0, lname + ".cv1", 1);
+    // chunk(2, 1)
+
+    nvinfer1::Dims d = cv1->getOutput(0)->getDimensions();
+    nvinfer1::ISliceLayer* split1 =
+            network->addSlice(*cv1->getOutput(0), nvinfer1::Dims3{0, 0, 0}, nvinfer1::Dims3{d.d[0] / 2, d.d[1], d.d[2]},
+                              nvinfer1::Dims3{1, 1, 1});
+    nvinfer1::ISliceLayer* split2 =
+            network->addSlice(*cv1->getOutput(0), nvinfer1::Dims3{d.d[0] / 2, 0, 0},
+                              nvinfer1::Dims3{d.d[0] / 2, d.d[1], d.d[2]}, nvinfer1::Dims3{1, 1, 1});
+    auto cv2 = convBnSiLU(network, weightMap, *split2->getOutput(0), c4, 3, 1, 1, lname + ".cv2", 1);
+
+    auto cv3 = convBnSiLU(network, weightMap, *cv2->getOutput(0), c4, 3, 1, 1, lname + ".cv3", 1);
+
+    ITensor* inputTensors[] = {split1->getOutput(0), split2->getOutput(0), cv2->getOutput(0), cv3->getOutput(0)};
+    auto cat = network->addConcatenation(inputTensors, 4);
+    auto cv4 = convBnSiLU(network, weightMap, *cat->getOutput(0), c2, 1, 1, 0, lname + ".cv4", 1);
+    return cv4;
+}
+
 ILayer* RepNCSPELAN4(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c1,
                     int c2, int c3, int c4, int c5, std::string lname) {

    auto cv1 = convBnSiLU(network, weightMap, input, c3, 1, 1, 0, lname + ".cv1", 1);
-    // 将cv1的输出分成两部分 chunk(2, 1)
+    // chunk(2, 1)

    nvinfer1::Dims d = cv1->getOutput(0)->getDimensions();
    nvinfer1::ISliceLayer* split1 =
@ -230,6 +253,14 @@ ILayer* RepNCSPELAN4(INetworkDefinition* network, std::map<std::string, Weights>
    return cv4;
 }

+ILayer* AConv(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c2,
+              std::string lname) {
+    auto pool = network->addPoolingNd(input, PoolingType::kAVERAGE, DimsHW{2, 2});
+    pool->setStrideNd(DimsHW{1, 1});
+    pool->setPaddingNd(DimsHW{0, 0});
+    auto cv1 = convBnSiLU(network, weightMap, *pool->getOutput(0), c2, 3, 2, 1, lname + ".cv1", 1);
+    return cv1;
+}
 ILayer* ADown(INetworkDefinition* network, std::map<std::string, Weights>& weightMap, ITensor& input, int c2,
              std::string lname) {
    int c_ = c2 / 2;
@ -426,7 +457,9 @@ std::vector<IConcatenationLayer*> DualDDetect(INetworkDefinition* network, std::
 std::vector<IConcatenationLayer*> DDetect(INetworkDefinition* network, std::map<std::string, Weights>& weightMap,
                                          std::vector<ILayer*> dets, int cls, std::vector<int> ch, std::string lname) {
    int c2 = std::max(int(ch[0] / 4), int(16 * 4));
-    int c3 = std::max(ch[0], std::min(cls * 2, 128));
+    //  max((ch[0], min((self.nc * 2, 128))))
+    // int c3 = std::max(ch[0], std::min(cls * 2, 128));
+    int c3 = std::max(ch[0], std::min(cls, 128));
    int reg_max = 16;

    std::vector<ILayer*> bboxlayers;
--- a/yolov9/src/model.cpp
+++ b/yolov9/src/model.cpp
@ -22,6 +22,723 @@ void Calibrator(IBuilder* builder, IBuilderConfig* config) {
 }
 #endif

+IHostMemory* build_engine_yolov9_t(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                   std::string& wts_name, bool isConvert) {
+    /* ------ Create the builder ------ */
+    INetworkDefinition* network = builder->createNetworkV2(0U);
+
+    ITensor* data = network->addInput(kInputTensorName, dt, Dims3{3, kInputH, kInputW});
+    assert(data);
+    std::map<std::string, Weights> weightMap = loadWeights(wts_name);
+
+    // # conv down
+    auto conv_1 = convBnSiLU(network, weightMap, *data, 16, 3, 2, 1, "model.0", 1);
+    // # conv down
+    auto conv_2 = convBnSiLU(network, weightMap, *conv_1->getOutput(0), 32, 3, 2, 1, "model.1");
+    // # elan-1 block
+    auto repncspelan_3 = ELAN1(network, weightMap, *conv_2->getOutput(0), 32, 32, 32, 16, "model.2");
+    // # avg-conv down
+    // [-1, 1, ADown, [256]],  # 4-P3/8
+    auto adown_4 = AConv(network, weightMap, *repncspelan_3->getOutput(0), 64, "model.3");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+    auto repncspelan_5 = RepNCSPELAN4(network, weightMap, *adown_4->getOutput(0), 64, 64, 64, 32, 3, "model.4");
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 6-P4/16
+    auto adown_6 = AConv(network, weightMap, *repncspelan_5->getOutput(0), 96, "model.5");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+    auto repncspelan_7 = RepNCSPELAN4(network, weightMap, *adown_6->getOutput(0), 96, 96, 96, 48, 3, "model.6");
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 8-P5/32
+    auto adown_8 = AConv(network, weightMap, *repncspelan_7->getOutput(0), 128, "model.7");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+    auto repncspelan_9 = RepNCSPELAN4(network, weightMap, *adown_8->getOutput(0), 128, 128, 128, 64, 3, "model.8");
+    // # elan-spp block
+    // [-1, 1, SPPELAN, [512, 256]],  # 10
+    auto sppelan_10 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 128, 128, 64, "model.9");
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_11 = network->addResize(*sppelan_10->getOutput(0));
+    upsample_11->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_11[] = {1.0, 2.0, 2.0};
+    upsample_11->setScales(scales_11, 3);
+    // [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+    ITensor* input_tensor_12[] = {upsample_11->getOutput(0), repncspelan_7->getOutput(0)};
+    auto cat_12 = network->addConcatenation(input_tensor_12, 2);
+
+    // # elan-2 block
+    auto repncspelan_13 = RepNCSPELAN4(network, weightMap, *cat_12->getOutput(0), 288, 96, 96, 48, 3, "model.12");
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_14 = network->addResize(*repncspelan_13->getOutput(0));
+    upsample_14->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_14[] = {1.0, 2.0, 2.0};
+    upsample_14->setScales(scales_14, 3);
+    // [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+    ITensor* input_tensor_15[] = {upsample_14->getOutput(0), repncspelan_5->getOutput(0)};
+    auto cat_15 = network->addConcatenation(input_tensor_15, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+    auto repncspelan_16 = RepNCSPELAN4(network, weightMap, *cat_15->getOutput(0), 192, 64, 64, 32, 3, "model.15");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [256]],
+    auto adown_17 = AConv(network, weightMap, *repncspelan_16->getOutput(0), 48, "model.16");
+    // [[-1, 13], 1, Concat, [1]],  # cat head P4
+    ITensor* input_tensor_18[] = {adown_17->getOutput(0), repncspelan_13->getOutput(0)};
+    auto cat_18 = network->addConcatenation(input_tensor_18, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+    auto repncspelan_19 = RepNCSPELAN4(network, weightMap, *cat_18->getOutput(0), 144, 96, 96, 48, 3, "model.18");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [512]],
+    auto adown_20 = AConv(network, weightMap, *repncspelan_19->getOutput(0), 64, "model.19");
+    // [[-1, 10], 1, Concat, [1]],  # cat head P5
+    ITensor* input_tensor_21[] = {adown_20->getOutput(0), sppelan_10->getOutput(0)};
+    auto cat_21 = network->addConcatenation(input_tensor_21, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+    auto repncspelan_22 = RepNCSPELAN4(network, weightMap, *cat_21->getOutput(0), 256, 128, 128, 64, 3, "model.21");
+
+    std::vector<IConcatenationLayer*> head;
+    if (!isConvert) {
+        // # elan-spp block
+        auto sppelan_23 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 128, 64, "model.22");
+
+        // # up-concat merge
+        auto upsample_24 = network->addResize(*sppelan_23->getOutput(0));
+        upsample_24->setResizeMode(ResizeMode::kNEAREST);
+        const float scales_24[] = {1.0, 2.0, 2.0};
+        upsample_24->setScales(scales_24, 3);
+        // [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+        ITensor* input_tensor_25[] = {upsample_24->getOutput(0), repncspelan_7->getOutput(0)};
+        auto cat_25 = network->addConcatenation(input_tensor_25, 2);
+
+        // # elan-2 block
+        auto repncspelan_26 = RepNCSPELAN4(network, weightMap, *cat_25->getOutput(0), 384, 96, 96, 48, 3, "model.25");
+
+        // # up-concat merge
+        auto upsample_27 = network->addResize(*repncspelan_26->getOutput(0));
+        upsample_27->setResizeMode(ResizeMode::kNEAREST);
+        const float scales_27[] = {1.0, 2.0, 2.0};
+        upsample_27->setScales(scales_27, 3);
+        // [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+        ITensor* input_tensor_28[] = {upsample_27->getOutput(0), repncspelan_5->getOutput(0)};
+        auto cat_28 = network->addConcatenation(input_tensor_28, 2);
+
+        // # elan-2 block
+        auto repncspelan_29 = RepNCSPELAN4(network, weightMap, *cat_28->getOutput(0), 256, 64, 64, 32, 3, "model.28");
+        head = DualDDetect(network, weightMap, std::vector<ILayer*>{repncspelan_16, repncspelan_19, repncspelan_22},
+                           kNumClass, {64, 96, 128}, "model.29");
+    } else {
+        head = DDetect(network, weightMap, std::vector<ILayer*>{repncspelan_16, repncspelan_19, repncspelan_22},
+                       kNumClass, {64, 96, 128}, "model.22");
+    }
+
+    nvinfer1::IPluginV2Layer* yolo = addYoLoLayer(network, head, false);
+    yolo->getOutput(0)->setName(kOutputTensorName);
+    network->markOutput(*yolo->getOutput(0));
+
+    builder->setMaxBatchSize(kBatchSize);
+    config->setMaxWorkspaceSize(16 * (1 << 20));
+
+#if defined(USE_FP16)
+    config->setFlag(nvinfer1::BuilderFlag::kFP16);
+#elif defined(USE_INT8)
+    std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
+    assert(builder->platformHasFastInt8());
+    config->setFlag(nvinfer1::BuilderFlag::kINT8);
+    auto* calibrator =
+            new Int8EntropyCalibrator2(1, kInputW, kInputH, gCalibTablePath, "int8calib.table", kInputTensorName);
+    config->setInt8Calibrator(calibrator);
+#endif
+
+    std::cout << "Building engine, please wait for a while..." << std::endl;
+    IHostMemory* serialized_model = builder->buildSerializedNetwork(*network, *config);
+    std::cout << "Build engine successfully!" << std::endl;
+
+    delete network;
+
+    // Release host memory
+    for (auto& mem : weightMap) {
+        free((void*)(mem.second.values));
+    }
+
+    return serialized_model;
+}
+
+IHostMemory* build_engine_yolov9_s(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                   std::string& wts_name, bool isConvert) {
+    /* ------ Create the builder ------ */
+    INetworkDefinition* network = builder->createNetworkV2(0U);
+
+    ITensor* data = network->addInput(kInputTensorName, dt, Dims3{3, kInputH, kInputW});
+    assert(data);
+    std::map<std::string, Weights> weightMap = loadWeights(wts_name);
+
+    // # conv down
+    auto conv_1 = convBnSiLU(network, weightMap, *data, 32, 3, 2, 1, "model.0", 1);
+    // # conv down
+    auto conv_2 = convBnSiLU(network, weightMap, *conv_1->getOutput(0), 64, 3, 2, 1, "model.1");
+    // # elan-1 block
+    auto repncspelan_3 = ELAN1(network, weightMap, *conv_2->getOutput(0), 32, 64, 64, 32, "model.2");
+    // # avg-conv down
+    auto adown_4 = AConv(network, weightMap, *repncspelan_3->getOutput(0), 128, "model.3");
+    // # elan-2 block
+    auto repncspelan_5 = RepNCSPELAN4(network, weightMap, *adown_4->getOutput(0), 128, 128, 128, 64, 3, "model.4");
+    // # avg-conv down
+    auto adown_6 = AConv(network, weightMap, *repncspelan_5->getOutput(0), 192, "model.5");
+    // # elan-2 block
+    auto repncspelan_7 = RepNCSPELAN4(network, weightMap, *adown_6->getOutput(0), 192, 192, 192, 96, 3, "model.6");
+    // # avg-conv down
+    auto adown_8 = AConv(network, weightMap, *repncspelan_7->getOutput(0), 256, "model.7");
+    // # elan-2 block
+    auto repncspelan_9 = RepNCSPELAN4(network, weightMap, *adown_8->getOutput(0), 256, 256, 256, 128, 3, "model.8");
+    // # elan-spp block
+    auto sppelan_10 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 256, 128, "model.9");
+
+    // # up-concat merge
+    auto upsample_11 = network->addResize(*sppelan_10->getOutput(0));
+    upsample_11->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_11[] = {1.0, 2.0, 2.0};
+    upsample_11->setScales(scales_11, 3);
+    // [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+    ITensor* input_tensor_12[] = {upsample_11->getOutput(0), repncspelan_7->getOutput(0)};
+    auto cat_12 = network->addConcatenation(input_tensor_12, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+    auto repncspelan_13 = RepNCSPELAN4(network, weightMap, *cat_12->getOutput(0), 192, 192, 192, 96, 3, "model.12");
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_14 = network->addResize(*repncspelan_13->getOutput(0));
+    upsample_14->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_14[] = {1.0, 2.0, 2.0};
+    upsample_14->setScales(scales_14, 3);
+    // [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+    ITensor* input_tensor_15[] = {upsample_14->getOutput(0), repncspelan_5->getOutput(0)};
+    auto cat_15 = network->addConcatenation(input_tensor_15, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+    auto repncspelan_16 = RepNCSPELAN4(network, weightMap, *cat_15->getOutput(0), 128, 128, 128, 64, 3, "model.15");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [256]],
+    auto adown_17 = AConv(network, weightMap, *repncspelan_16->getOutput(0), 96, "model.16");
+    // [[-1, 13], 1, Concat, [1]],  # cat head P4
+    ITensor* input_tensor_18[] = {adown_17->getOutput(0), repncspelan_13->getOutput(0)};
+    auto cat_18 = network->addConcatenation(input_tensor_18, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+    auto repncspelan_19 = RepNCSPELAN4(network, weightMap, *cat_18->getOutput(0), 768, 192, 192, 96, 3, "model.18");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [512]],
+    auto adown_20 = AConv(network, weightMap, *repncspelan_19->getOutput(0), 128, "model.19");
+    // [[-1, 10], 1, Concat, [1]],  # cat head P5
+    ITensor* input_tensor_21[] = {adown_20->getOutput(0), sppelan_10->getOutput(0)};
+    auto cat_21 = network->addConcatenation(input_tensor_21, 2);
+
+    // # elan-2 block
+    auto repncspelan_22 = RepNCSPELAN4(network, weightMap, *cat_21->getOutput(0), 1024, 256, 256, 128, 1, "model.21");
+    std::vector<IConcatenationLayer*> head;
+    if (!isConvert) {
+        // # elan-spp block
+        auto sppelan_23 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 256, 128, "model.22");
+
+        // # up-concat merge
+        auto upsample_24 = network->addResize(*sppelan_23->getOutput(0));
+        upsample_24->setResizeMode(ResizeMode::kNEAREST);
+        const float scales_24[] = {1.0, 2.0, 2.0};
+        upsample_24->setScales(scales_24, 3);
+        // [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+        ITensor* input_tensor_25[] = {upsample_24->getOutput(0), repncspelan_7->getOutput(0)};
+        auto cat_25 = network->addConcatenation(input_tensor_25, 2);
+
+        // # elan-2 block
+        auto repncspelan_26 = RepNCSPELAN4(network, weightMap, *cat_25->getOutput(0), 384, 192, 192, 96, 3, "model.25");
+
+        // # up-concat merge
+        auto upsample_27 = network->addResize(*repncspelan_26->getOutput(0));
+        upsample_27->setResizeMode(ResizeMode::kNEAREST);
+        const float scales_27[] = {1.0, 2.0, 2.0};
+        upsample_27->setScales(scales_27, 3);
+        // [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+        ITensor* input_tensor_28[] = {upsample_27->getOutput(0), repncspelan_5->getOutput(0)};
+        auto cat_28 = network->addConcatenation(input_tensor_28, 2);
+
+        // # elan-2 block
+        auto repncspelan_29 = RepNCSPELAN4(network, weightMap, *cat_28->getOutput(0), 256, 128, 128, 64, 3, "model.28");
+        head = DualDDetect(network, weightMap, std::vector<ILayer*>{repncspelan_16, repncspelan_19, repncspelan_22},
+                           kNumClass, {128, 192, 256}, "model.29");
+    } else {
+        head = DDetect(network, weightMap, std::vector<ILayer*>{repncspelan_16, repncspelan_19, repncspelan_22},
+                       kNumClass, {128, 192, 256}, "model.22");
+    }
+
+    nvinfer1::IPluginV2Layer* yolo = addYoLoLayer(network, head, false);
+    yolo->getOutput(0)->setName(kOutputTensorName);
+    network->markOutput(*yolo->getOutput(0));
+
+    builder->setMaxBatchSize(kBatchSize);
+    config->setMaxWorkspaceSize(16 * (1 << 20));
+
+#if defined(USE_FP16)
+    config->setFlag(nvinfer1::BuilderFlag::kFP16);
+#elif defined(USE_INT8)
+    std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
+    assert(builder->platformHasFastInt8());
+    config->setFlag(nvinfer1::BuilderFlag::kINT8);
+    auto* calibrator =
+            new Int8EntropyCalibrator2(1, kInputW, kInputH, gCalibTablePath, "int8calib.table", kInputTensorName);
+    config->setInt8Calibrator(calibrator);
+#endif
+
+    std::cout << "Building engine, please wait for a while..." << std::endl;
+    IHostMemory* serialized_model = builder->buildSerializedNetwork(*network, *config);
+    std::cout << "Build engine successfully!" << std::endl;
+
+    delete network;
+
+    // Release host memory
+    for (auto& mem : weightMap) {
+        free((void*)(mem.second.values));
+    }
+
+    return serialized_model;
+}
+IHostMemory* build_engine_yolov9_m(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                   std::string& wts_name, bool isConvert) {
+    /* ------ Create the builder ------ */
+    INetworkDefinition* network = builder->createNetworkV2(0U);
+
+    ITensor* data = network->addInput(kInputTensorName, dt, Dims3{3, kInputH, kInputW});
+    assert(data);
+    std::map<std::string, Weights> weightMap = loadWeights(wts_name);
+
+    int begin = isConvert ? 0 : 1;
+
+    // # conv down
+    // [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+    auto conv_1 = convBnSiLU(network, weightMap, *data, 32, 3, 2, 1, "model." + std::to_string(begin), 1);
+    begin += 1;
+    // # conv down
+    // [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+    auto conv_2 = convBnSiLU(network, weightMap, *conv_1->getOutput(0), 64, 3, 2, 1, "model." + std::to_string(begin));
+    begin += 1;
+    // # elan-1 block
+    // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
+    auto repncspelan_3 = RepNCSPELAN4(network, weightMap, *conv_2->getOutput(0), 128, 128, 128, 64, 1,
+                                      "model." + std::to_string(begin));
+    begin += 1;
+    // # avg-conv down
+    // [-1, 1, ADown, [256]],  # 4-P3/8
+    auto adown_4 = AConv(network, weightMap, *repncspelan_3->getOutput(0), 240, "model." + std::to_string(begin));
+    begin += 1;
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+    auto repncspelan_5 = RepNCSPELAN4(network, weightMap, *adown_4->getOutput(0), 256, 240, 240, 120, 1,
+                                      "model." + std::to_string(begin));
+    begin += 1;
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 6-P4/16
+    auto adown_6 = AConv(network, weightMap, *repncspelan_5->getOutput(0), 360, "model." + std::to_string(begin));
+    begin += 1;
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+    auto repncspelan_7 = RepNCSPELAN4(network, weightMap, *adown_6->getOutput(0), 512, 360, 360, 180, 1,
+                                      "model." + std::to_string(begin));
+    begin += 1;
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 8-P5/32
+    auto adown_8 = AConv(network, weightMap, *repncspelan_7->getOutput(0), 480, "model." + std::to_string(begin));
+    begin += 1;
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+    auto repncspelan_9 = RepNCSPELAN4(network, weightMap, *adown_8->getOutput(0), 512, 480, 480, 240, 1,
+                                      "model." + std::to_string(begin));
+    begin += 1;
+    // # elan-spp block
+    // [-1, 1, SPPELAN, [512, 256]],  # 10
+    auto sppelan_10 =
+            SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 480, 240, "model." + std::to_string(begin));
+    begin += 3;
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_11 = network->addResize(*sppelan_10->getOutput(0));
+    upsample_11->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_11[] = {1.0, 2.0, 2.0};
+    upsample_11->setScales(scales_11, 3);
+    // [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+    ITensor* input_tensor_12[] = {upsample_11->getOutput(0), repncspelan_7->getOutput(0)};
+    auto cat_12 = network->addConcatenation(input_tensor_12, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+    auto repncspelan_13 = RepNCSPELAN4(network, weightMap, *cat_12->getOutput(0), 1536, 360, 360, 180, 1,
+                                       "model." + std::to_string(begin));
+    begin += 3;
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_14 = network->addResize(*repncspelan_13->getOutput(0));
+    upsample_14->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_14[] = {1.0, 2.0, 2.0};
+    upsample_14->setScales(scales_14, 3);
+    // [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+    ITensor* input_tensor_15[] = {upsample_14->getOutput(0), repncspelan_5->getOutput(0)};
+    auto cat_15 = network->addConcatenation(input_tensor_15, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+    auto repncspelan_16 = RepNCSPELAN4(network, weightMap, *cat_15->getOutput(0), 1024, 240, 240, 120, 1,
+                                       "model." + std::to_string(begin));
+    begin += 1;
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [256]],
+    auto adown_17 = AConv(network, weightMap, *repncspelan_16->getOutput(0), 184, "model." + std::to_string(begin));
+    begin += 2;
+    // [[-1, 13], 1, Concat, [1]],  # cat head P4
+    ITensor* input_tensor_18[] = {adown_17->getOutput(0), repncspelan_13->getOutput(0)};
+    auto cat_18 = network->addConcatenation(input_tensor_18, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+    auto repncspelan_19 = RepNCSPELAN4(network, weightMap, *cat_18->getOutput(0), 768, 360, 360, 180, 1,
+                                       "model." + std::to_string(begin));
+    begin += 1;
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [512]],
+    auto adown_20 = AConv(network, weightMap, *repncspelan_19->getOutput(0), 240, "model." + std::to_string(begin));
+    begin += 2;
+    // [[-1, 10], 1, Concat, [1]],  # cat head P5
+    ITensor* input_tensor_21[] = {adown_20->getOutput(0), sppelan_10->getOutput(0)};
+    auto cat_21 = network->addConcatenation(input_tensor_21, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+    auto repncspelan_22 = RepNCSPELAN4(network, weightMap, *cat_21->getOutput(0), 1024, 480, 480, 240, 1,
+                                       "model." + std::to_string(begin));
+    begin += 1;
+    std::vector<IConcatenationLayer*> head;
+    if (!isConvert) {
+        // # routing
+        // [5, 1, CBLinear, [[256]]], # 23
+        auto cblinear_23 = CBLinear(network, weightMap, *repncspelan_5->getOutput(0), {240}, 1, 1, 0, 1,
+                                    "model." + std::to_string(begin));
+        begin += 1;
+        // [7, 1, CBLinear, [[256, 512]]], # 24
+        auto cblinear_24 = CBLinear(network, weightMap, *repncspelan_7->getOutput(0), {240, 360}, 1, 1, 0, 1,
+                                    "model." + std::to_string(begin));
+        begin += 1;
+        // [9, 1, CBLinear, [[256, 512, 512]]], # 25
+        auto cblinear_25 = CBLinear(network, weightMap, *repncspelan_9->getOutput(0), {240, 360, 480}, 1, 1, 0, 1,
+                                    "model." + std::to_string(begin));
+        begin += 1;
+
+        // # conv down
+        // [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
+        auto conv_26 = convBnSiLU(network, weightMap, *data, 32, 3, 2, 1, "model." + std::to_string(begin), 1);
+        begin += 1;
+
+        // # conv down
+        // [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
+        auto conv_27 =
+                convBnSiLU(network, weightMap, *conv_26->getOutput(0), 64, 3, 2, 1, "model." + std::to_string(begin));
+        begin += 1;
+
+        // # elan-1 block
+        // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
+        auto repncspelan_28 = RepNCSPELAN4(network, weightMap, *conv_27->getOutput(0), 128, 128, 128, 64, 1,
+                                           "model." + std::to_string(begin));
+        begin += 1;
+
+        // # avg-conv down fuse
+        // [-1, 1, ADown, [256]],  # 29-P3/8
+        auto adown_29 = AConv(network, weightMap, *repncspelan_28->getOutput(0), 240, "model." + std::to_string(begin));
+        begin += 2;
+        // [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30
+        auto cbfuse = CBFuse(network, {cblinear_23, cblinear_24, cblinear_25, std::vector<ILayer*>{adown_29}},
+                             {0, 0, 0, 0}, {8, 16, 32, 8});
+
+        // # elan-2 block
+        // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
+        auto repncspelan_31 = RepNCSPELAN4(network, weightMap, *cbfuse->getOutput(0), 256, 240, 240, 120, 1,
+                                           "model." + std::to_string(begin));
+        begin += 1;
+
+        // # avg-conv down fuse
+        // [-1, 1, ADown, [512]],  # 32-P4/16
+        auto adown_32 = AConv(network, weightMap, *repncspelan_31->getOutput(0), 360, "model." + std::to_string(begin));
+        begin += 2;
+        // [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33
+        auto cbfuse_33 =
+                CBFuse(network, {cblinear_24, cblinear_25, std::vector<ILayer*>{adown_32}}, {1, 1, 0}, {16, 32, 16});
+
+        // # elan-2 block
+        // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
+        auto repncspelan_34 = RepNCSPELAN4(network, weightMap, *cbfuse_33->getOutput(0), 512, 360, 360, 180, 1,
+                                           "model." + std::to_string(begin));
+        begin += 1;
+
+        // # avg-conv down fuse
+        // [-1, 1, ADown, [512]],  # 35-P5/32
+        auto adown_35 = AConv(network, weightMap, *repncspelan_34->getOutput(0), 480, "model." + std::to_string(begin));
+        begin += 2;
+
+        // [[25, -1], 1, CBFuse, [[2]]], # 36
+        auto cbfuse_36 = CBFuse(network, {cblinear_25, std::vector<ILayer*>{adown_35}}, {2, 0}, {32, 32});
+
+        // # elan-2 block
+        // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
+        auto repncspelan_37 = RepNCSPELAN4(network, weightMap, *cbfuse_36->getOutput(0), 512, 480, 480, 240, 1,
+                                           "model." + std::to_string(begin));
+        begin += 1;
+
+        // # detection head
+        // # detect
+        // [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # DualDDetect(A3, A4, A5, P3, P4, P5)
+        head = DualDDetect(network, weightMap, std::vector<ILayer*>{repncspelan_31, repncspelan_34, repncspelan_37},
+                           kNumClass, {240, 360, 480}, "model." + std::to_string(begin));
+    } else {
+        // # detection head
+        // # detect
+        // [[16, 19, 22], 1, DDetect, [nc]],  # DDetect(P3, P4, P5)
+        head = DDetect(network, weightMap, std::vector<ILayer*>{repncspelan_16, repncspelan_19, repncspelan_22},
+                       kNumClass, {240, 360, 480}, "model." + std::to_string(begin));
+    }
+
+    nvinfer1::IPluginV2Layer* yolo = addYoLoLayer(network, head, false);
+    yolo->getOutput(0)->setName(kOutputTensorName);
+    network->markOutput(*yolo->getOutput(0));
+
+    builder->setMaxBatchSize(kBatchSize);
+    config->setMaxWorkspaceSize(16 * (1 << 20));
+
+#if defined(USE_FP16)
+    config->setFlag(nvinfer1::BuilderFlag::kFP16);
+#elif defined(USE_INT8)
+    std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
+    assert(builder->platformHasFastInt8());
+    config->setFlag(nvinfer1::BuilderFlag::kINT8);
+    auto* calibrator =
+            new Int8EntropyCalibrator2(1, kInputW, kInputH, gCalibTablePath, "int8calib.table", kInputTensorName);
+    config->setInt8Calibrator(calibrator);
+#endif
+
+    std::cout << "Building engine, please wait for a while..." << std::endl;
+    IHostMemory* serialized_model = builder->buildSerializedNetwork(*network, *config);
+    std::cout << "Build engine successfully!" << std::endl;
+
+    delete network;
+
+    // Release host memory
+    for (auto& mem : weightMap) {
+        free((void*)(mem.second.values));
+    }
+
+    return serialized_model;
+}
+
+IHostMemory* build_engine_yolov9_c(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                   std::string& wts_name) {
+    /* ------ Create the builder ------ */
+    INetworkDefinition* network = builder->createNetworkV2(0U);
+
+    ITensor* data = network->addInput(kInputTensorName, dt, Dims3{3, kInputH, kInputW});
+    assert(data);
+    std::map<std::string, Weights> weightMap = loadWeights(wts_name);
+
+    // # conv down
+    // [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+    auto conv_1 = convBnSiLU(network, weightMap, *data, 64, 3, 2, 1, "model.1", 1);
+    // # conv down
+    // [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+    auto conv_2 = convBnSiLU(network, weightMap, *conv_1->getOutput(0), 128, 3, 2, 1, "model.2");
+    // # elan-1 block
+    // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
+    auto repncspelan_3 = RepNCSPELAN4(network, weightMap, *conv_2->getOutput(0), 128, 256, 128, 64, 1, "model.3");
+    // # avg-conv down
+    // [-1, 1, ADown, [256]],  # 4-P3/8
+    auto adown_4 = ADown(network, weightMap, *repncspelan_3->getOutput(0), 256, "model.4");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+    auto repncspelan_5 = RepNCSPELAN4(network, weightMap, *adown_4->getOutput(0), 256, 512, 256, 128, 1, "model.5");
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 6-P4/16
+    auto adown_6 = ADown(network, weightMap, *repncspelan_5->getOutput(0), 512, "model.6");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+    auto repncspelan_7 = RepNCSPELAN4(network, weightMap, *adown_6->getOutput(0), 512, 512, 512, 256, 1, "model.7");
+    // # avg-conv down
+    // [-1, 1, ADown, [512]],  # 8-P5/32
+    auto adown_8 = ADown(network, weightMap, *repncspelan_7->getOutput(0), 512, "model.8");
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+    auto repncspelan_9 = RepNCSPELAN4(network, weightMap, *adown_8->getOutput(0), 512, 512, 512, 256, 1, "model.9");
+    // # elan-spp block
+    // [-1, 1, SPPELAN, [512, 256]],  # 10
+    auto sppelan_10 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 512, 256, "model.10");
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_11 = network->addResize(*sppelan_10->getOutput(0));
+    upsample_11->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_11[] = {1.0, 2.0, 2.0};
+    upsample_11->setScales(scales_11, 3);
+    // [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+    ITensor* input_tensor_12[] = {upsample_11->getOutput(0), repncspelan_7->getOutput(0)};
+    auto cat_12 = network->addConcatenation(input_tensor_12, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+    auto repncspelan_13 = RepNCSPELAN4(network, weightMap, *cat_12->getOutput(0), 1536, 512, 512, 256, 1, "model.13");
+
+    // # up-concat merge
+    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+    auto upsample_14 = network->addResize(*repncspelan_13->getOutput(0));
+    upsample_14->setResizeMode(ResizeMode::kNEAREST);
+    const float scales_14[] = {1.0, 2.0, 2.0};
+    upsample_14->setScales(scales_14, 3);
+    // [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+    ITensor* input_tensor_15[] = {upsample_14->getOutput(0), repncspelan_5->getOutput(0)};
+    auto cat_15 = network->addConcatenation(input_tensor_15, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+    auto repncspelan_16 = RepNCSPELAN4(network, weightMap, *cat_15->getOutput(0), 1024, 256, 256, 128, 1, "model.16");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [256]],
+    auto adown_17 = ADown(network, weightMap, *repncspelan_16->getOutput(0), 256, "model.17");
+    // [[-1, 13], 1, Concat, [1]],  # cat head P4
+    ITensor* input_tensor_18[] = {adown_17->getOutput(0), repncspelan_13->getOutput(0)};
+    auto cat_18 = network->addConcatenation(input_tensor_18, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+    auto repncspelan_19 = RepNCSPELAN4(network, weightMap, *cat_18->getOutput(0), 768, 512, 512, 256, 1, "model.19");
+
+    // # avg-conv-down merge
+    // [-1, 1, ADown, [512]],
+    auto adown_20 = ADown(network, weightMap, *repncspelan_19->getOutput(0), 512, "model.20");
+    // [[-1, 10], 1, Concat, [1]],  # cat head P5
+    ITensor* input_tensor_21[] = {adown_20->getOutput(0), sppelan_10->getOutput(0)};
+    auto cat_21 = network->addConcatenation(input_tensor_21, 2);
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+    auto repncspelan_22 = RepNCSPELAN4(network, weightMap, *cat_21->getOutput(0), 1024, 512, 512, 256, 1, "model.22");
+
+    // # multi-level reversible auxiliary branch
+
+    // # routing
+    // [5, 1, CBLinear, [[256]]], # 23
+    auto cblinear_23 = CBLinear(network, weightMap, *repncspelan_5->getOutput(0), {256}, 1, 1, 0, 1, "model.23");
+    // [7, 1, CBLinear, [[256, 512]]], # 24
+    auto cblinear_24 = CBLinear(network, weightMap, *repncspelan_7->getOutput(0), {256, 512}, 1, 1, 0, 1, "model.24");
+    // [9, 1, CBLinear, [[256, 512, 512]]], # 25
+    auto cblinear_25 =
+            CBLinear(network, weightMap, *repncspelan_9->getOutput(0), {256, 512, 512}, 1, 1, 0, 1, "model.25");
+
+    // # conv down
+    // [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
+    auto conv_26 = convBnSiLU(network, weightMap, *data, 64, 3, 2, 1, "model.26", 1);
+
+    // # conv down
+    // [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
+    auto conv_27 = convBnSiLU(network, weightMap, *conv_26->getOutput(0), 128, 3, 2, 1, "model.27");
+
+    // # elan-1 block
+    // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
+    auto repncspelan_28 = RepNCSPELAN4(network, weightMap, *conv_27->getOutput(0), 128, 256, 128, 64, 1, "model.28");
+
+    // # avg-conv down fuse
+    // [-1, 1, ADown, [256]],  # 29-P3/8
+    auto adown_29 = ADown(network, weightMap, *repncspelan_28->getOutput(0), 256, "model.29");
+    // [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30
+    auto cbfuse = CBFuse(network, {cblinear_23, cblinear_24, cblinear_25, std::vector<ILayer*>{adown_29}}, {0, 0, 0, 0},
+                         {8, 16, 32, 8});
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
+    auto repncspelan_31 = RepNCSPELAN4(network, weightMap, *cbfuse->getOutput(0), 256, 512, 256, 128, 1, "model.31");
+
+    // # avg-conv down fuse
+    // [-1, 1, ADown, [512]],  # 32-P4/16
+    auto adown_32 = ADown(network, weightMap, *repncspelan_31->getOutput(0), 512, "model.32");
+    // [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33
+    auto cbfuse_33 =
+            CBFuse(network, {cblinear_24, cblinear_25, std::vector<ILayer*>{adown_32}}, {1, 1, 0}, {16, 32, 16});
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
+    auto repncspelan_34 = RepNCSPELAN4(network, weightMap, *cbfuse_33->getOutput(0), 512, 512, 512, 256, 1, "model.34");
+
+    // # avg-conv down fuse
+    // [-1, 1, ADown, [512]],  # 35-P5/32
+    auto adown_35 = ADown(network, weightMap, *repncspelan_34->getOutput(0), 512, "model.35");
+
+    // [[25, -1], 1, CBFuse, [[2]]], # 36
+    auto cbfuse_36 = CBFuse(network, {cblinear_25, std::vector<ILayer*>{adown_35}}, {2, 0}, {32, 32});
+
+    // # elan-2 block
+    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
+    auto repncspelan_37 = RepNCSPELAN4(network, weightMap, *cbfuse_36->getOutput(0), 512, 512, 512, 256, 1, "model.37");
+
+    // # detection head
+    // # detect
+    // [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # DualDDetect(A3, A4, A5, P3, P4, P5)
+    auto dualddetect_38 =
+            DualDDetect(network, weightMap, std::vector<ILayer*>{repncspelan_31, repncspelan_34, repncspelan_37},
+                        kNumClass, {512, 512, 512}, "model.38");
+
+    nvinfer1::IPluginV2Layer* yolo = addYoLoLayer(network, dualddetect_38, false);
+    yolo->getOutput(0)->setName(kOutputTensorName);
+    network->markOutput(*yolo->getOutput(0));
+
+    builder->setMaxBatchSize(kBatchSize);
+    config->setMaxWorkspaceSize(16 * (1 << 20));
+
+#if defined(USE_FP16)
+    config->setFlag(nvinfer1::BuilderFlag::kFP16);
+#elif defined(USE_INT8)
+    std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
+    assert(builder->platformHasFastInt8());
+    config->setFlag(nvinfer1::BuilderFlag::kINT8);
+    auto* calibrator =
+            new Int8EntropyCalibrator2(1, kInputW, kInputH, gCalibTablePath, "int8calib.table", kInputTensorName);
+    config->setInt8Calibrator(calibrator);
+#endif
+
+    std::cout << "Building engine, please wait for a while..." << std::endl;
+    IHostMemory* serialized_model = builder->buildSerializedNetwork(*network, *config);
+    std::cout << "Build engine successfully!" << std::endl;
+
+    delete network;
+
+    // Release host memory
+    for (auto& mem : weightMap) {
+        free((void*)(mem.second.values));
+    }
+
+    return serialized_model;
+}
+
 IHostMemory* build_engine_yolov9_e(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
                                   std::string& wts_name) {
    /* ------ Create the builder ------ */
@ -247,194 +964,9 @@ IHostMemory* build_engine_yolov9_e(unsigned int maxBatchSize, IBuilder* builder,

    return serialized_model;
 }
-IHostMemory* build_engine_yolov9_c(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
-                                   std::string& wts_name) {
-    /* ------ Create the builder ------ */
-    INetworkDefinition* network = builder->createNetworkV2(0U);

-    ITensor* data = network->addInput(kInputTensorName, dt, Dims3{3, kInputH, kInputW});
-    assert(data);
-    std::map<std::string, Weights> weightMap = loadWeights(wts_name);
-
-    // # conv down
-    // [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
-    auto conv_1 = convBnSiLU(network, weightMap, *data, 64, 3, 2, 1, "model.1", 1);
-    // # conv down
-    // [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
-    auto conv_2 = convBnSiLU(network, weightMap, *conv_1->getOutput(0), 128, 3, 2, 1, "model.2");
-    // # elan-1 block
-    // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
-    auto repncspelan_3 = RepNCSPELAN4(network, weightMap, *conv_2->getOutput(0), 128, 256, 128, 64, 1, "model.3");
-    // # avg-conv down
-    // [-1, 1, ADown, [256]],  # 4-P3/8
-    auto adown_4 = ADown(network, weightMap, *repncspelan_3->getOutput(0), 256, "model.4");
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
-    auto repncspelan_5 = RepNCSPELAN4(network, weightMap, *adown_4->getOutput(0), 256, 512, 256, 128, 1, "model.5");
-    // # avg-conv down
-    // [-1, 1, ADown, [512]],  # 6-P4/16
-    auto adown_6 = ADown(network, weightMap, *repncspelan_5->getOutput(0), 512, "model.6");
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
-    auto repncspelan_7 = RepNCSPELAN4(network, weightMap, *adown_6->getOutput(0), 512, 512, 512, 256, 1, "model.7");
-    // # avg-conv down
-    // [-1, 1, ADown, [512]],  # 8-P5/32
-    auto adown_8 = ADown(network, weightMap, *repncspelan_7->getOutput(0), 512, "model.8");
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
-    auto repncspelan_9 = RepNCSPELAN4(network, weightMap, *adown_8->getOutput(0), 512, 512, 512, 256, 1, "model.9");
-    // # elan-spp block
-    // [-1, 1, SPPELAN, [512, 256]],  # 10
-    auto sppelan_10 = SPPELAN(network, weightMap, *repncspelan_9->getOutput(0), 512, 512, 256, "model.10");
-
-    // # up-concat merge
-    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
-    auto upsample_11 = network->addResize(*sppelan_10->getOutput(0));
-    upsample_11->setResizeMode(ResizeMode::kNEAREST);
-    const float scales_11[] = {1.0, 2.0, 2.0};
-    upsample_11->setScales(scales_11, 3);
-    // [[-1, 7], 1, Concat, [1]],  # cat backbone P4
-    ITensor* input_tensor_12[] = {upsample_11->getOutput(0), repncspelan_7->getOutput(0)};
-    auto cat_12 = network->addConcatenation(input_tensor_12, 2);
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
-    auto repncspelan_13 = RepNCSPELAN4(network, weightMap, *cat_12->getOutput(0), 1536, 512, 512, 256, 1, "model.13");
-
-    // # up-concat merge
-    // [-1, 1, nn.Upsample, [None, 2, 'nearest']],
-    auto upsample_14 = network->addResize(*repncspelan_13->getOutput(0));
-    upsample_14->setResizeMode(ResizeMode::kNEAREST);
-    const float scales_14[] = {1.0, 2.0, 2.0};
-    upsample_14->setScales(scales_14, 3);
-    // [[-1, 5], 1, Concat, [1]],  # cat backbone P3
-    ITensor* input_tensor_15[] = {upsample_14->getOutput(0), repncspelan_5->getOutput(0)};
-    auto cat_15 = network->addConcatenation(input_tensor_15, 2);
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
-    auto repncspelan_16 = RepNCSPELAN4(network, weightMap, *cat_15->getOutput(0), 1024, 256, 256, 128, 1, "model.16");
-
-    // # avg-conv-down merge
-    // [-1, 1, ADown, [256]],
-    auto adown_17 = ADown(network, weightMap, *repncspelan_16->getOutput(0), 256, "model.17");
-    // [[-1, 13], 1, Concat, [1]],  # cat head P4
-    ITensor* input_tensor_18[] = {adown_17->getOutput(0), repncspelan_13->getOutput(0)};
-    auto cat_18 = network->addConcatenation(input_tensor_18, 2);
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
-    auto repncspelan_19 = RepNCSPELAN4(network, weightMap, *cat_18->getOutput(0), 768, 512, 512, 256, 1, "model.19");
-
-    // # avg-conv-down merge
-    // [-1, 1, ADown, [512]],
-    auto adown_20 = ADown(network, weightMap, *repncspelan_19->getOutput(0), 512, "model.20");
-    // [[-1, 10], 1, Concat, [1]],  # cat head P5
-    ITensor* input_tensor_21[] = {adown_20->getOutput(0), sppelan_10->getOutput(0)};
-    auto cat_21 = network->addConcatenation(input_tensor_21, 2);
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
-    auto repncspelan_22 = RepNCSPELAN4(network, weightMap, *cat_21->getOutput(0), 1024, 512, 512, 256, 1, "model.22");
-
-    // # multi-level reversible auxiliary branch
-
-    // # routing
-    // [5, 1, CBLinear, [[256]]], # 23
-    auto cblinear_23 = CBLinear(network, weightMap, *repncspelan_5->getOutput(0), {256}, 1, 1, 0, 1, "model.23");
-    // [7, 1, CBLinear, [[256, 512]]], # 24
-    auto cblinear_24 = CBLinear(network, weightMap, *repncspelan_7->getOutput(0), {256, 512}, 1, 1, 0, 1, "model.24");
-    // [9, 1, CBLinear, [[256, 512, 512]]], # 25
-    auto cblinear_25 =
-            CBLinear(network, weightMap, *repncspelan_9->getOutput(0), {256, 512, 512}, 1, 1, 0, 1, "model.25");
-
-    // # conv down
-    // [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
-    auto conv_26 = convBnSiLU(network, weightMap, *data, 64, 3, 2, 1, "model.26", 1);
-
-    // # conv down
-    // [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
-    auto conv_27 = convBnSiLU(network, weightMap, *conv_26->getOutput(0), 128, 3, 2, 1, "model.27");
-
-    // # elan-1 block
-    // [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
-    auto repncspelan_28 = RepNCSPELAN4(network, weightMap, *conv_27->getOutput(0), 128, 256, 128, 64, 1, "model.28");
-
-    // # avg-conv down fuse
-    // [-1, 1, ADown, [256]],  # 29-P3/8
-    auto adown_29 = ADown(network, weightMap, *repncspelan_28->getOutput(0), 256, "model.29");
-    // [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30
-    auto cbfuse = CBFuse(network, {cblinear_23, cblinear_24, cblinear_25, std::vector<ILayer*>{adown_29}}, {0, 0, 0, 0},
-                         {8, 16, 32, 8});
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
-    auto repncspelan_31 = RepNCSPELAN4(network, weightMap, *cbfuse->getOutput(0), 256, 512, 256, 128, 1, "model.31");
-
-    // # avg-conv down fuse
-    // [-1, 1, ADown, [512]],  # 32-P4/16
-    auto adown_32 = ADown(network, weightMap, *repncspelan_31->getOutput(0), 512, "model.32");
-    // [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33
-    auto cbfuse_33 =
-            CBFuse(network, {cblinear_24, cblinear_25, std::vector<ILayer*>{adown_32}}, {1, 1, 0}, {16, 32, 16});
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
-    auto repncspelan_34 = RepNCSPELAN4(network, weightMap, *cbfuse_33->getOutput(0), 512, 512, 512, 256, 1, "model.34");
-
-    // # avg-conv down fuse
-    // [-1, 1, ADown, [512]],  # 35-P5/32
-    auto adown_35 = ADown(network, weightMap, *repncspelan_34->getOutput(0), 512, "model.35");
-
-    // [[25, -1], 1, CBFuse, [[2]]], # 36
-    auto cbfuse_36 = CBFuse(network, {cblinear_25, std::vector<ILayer*>{adown_35}}, {2, 0}, {32, 32});
-
-    // # elan-2 block
-    // [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
-    auto repncspelan_37 = RepNCSPELAN4(network, weightMap, *cbfuse_36->getOutput(0), 512, 512, 512, 256, 1, "model.37");
-
-    // # detection head
-    // # detect
-    // [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # DualDDetect(A3, A4, A5, P3, P4, P5)
-    auto dualddetect_38 =
-            DualDDetect(network, weightMap, std::vector<ILayer*>{repncspelan_31, repncspelan_34, repncspelan_37},
-                        kNumClass, {512, 512, 512}, "model.38");
-
-    nvinfer1::IPluginV2Layer* yolo = addYoLoLayer(network, dualddetect_38, false);
-    yolo->getOutput(0)->setName(kOutputTensorName);
-    network->markOutput(*yolo->getOutput(0));
-
-    builder->setMaxBatchSize(kBatchSize);
-    config->setMaxWorkspaceSize(16 * (1 << 20));
-
-#if defined(USE_FP16)
-    config->setFlag(nvinfer1::BuilderFlag::kFP16);
-#elif defined(USE_INT8)
-    std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;
-    assert(builder->platformHasFastInt8());
-    config->setFlag(nvinfer1::BuilderFlag::kINT8);
-    auto* calibrator =
-            new Int8EntropyCalibrator2(1, kInputW, kInputH, gCalibTablePath, "int8calib.table", kInputTensorName);
-    config->setInt8Calibrator(calibrator);
-#endif
-
-    std::cout << "Building engine, please wait for a while..." << std::endl;
-    IHostMemory* serialized_model = builder->buildSerializedNetwork(*network, *config);
-    std::cout << "Build engine successfully!" << std::endl;
-
-    delete network;
-
-    // Release host memory
-    for (auto& mem : weightMap) {
-        free((void*)(mem.second.values));
-    }
-
-    return serialized_model;
-}
-
-nvinfer1::IHostMemory* build_engine_gelan_e(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
-                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
-                                            std::string& wts_name) {
+IHostMemory* build_engine_gelan_e(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                  std::string& wts_name) {
    /* ------ Create the builder ------ */
    INetworkDefinition* network = builder->createNetworkV2(0U);

@ -625,9 +1157,8 @@ nvinfer1::IHostMemory* build_engine_gelan_e(unsigned int maxBatchSize, nvinfer1:

    return serialized_model;
 }
-nvinfer1::IHostMemory* build_engine_gelan_c(unsigned int maxBatchSize, nvinfer1::IBuilder* builder,
-                                            nvinfer1::IBuilderConfig* config, nvinfer1::DataType dt,
-                                            std::string& wts_name) {
+IHostMemory* build_engine_gelan_c(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt,
+                                  std::string& wts_name) {
    /* ------ Create the builder ------ */
    INetworkDefinition* network = builder->createNetworkV2(0U);