yolov8x-p2 实现 tensorrt 推理

J ..2023-10-24 3:05

简述

在最开始的yolov8提供的不同size的版本，包括n、s、m、l、x（模型规模依次增大，通过depth, width, max_channels控制大小），这些都是通过P3、P4和P5提取图片特征；

正常的yolov8对象检测模型输出层是P3、P4、P5三个输出层，为了提升对小目标的检测能力，新版本的yolov8 已经包含了P2层（P2层做的卷积次数少，特征图的尺寸（分辨率）较大，更加利于小目标识别），有四个输出层。Backbone部分的结果没有改变，但是Neck跟Head部分模型结构做了调整。

yolov8-p2 yaml

cpp 复制代码 
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P2-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 80  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]
  s: [0.33, 0.50, 1024]
  m: [0.67, 0.75, 768]
  l: [1.00, 1.00, 512]
  x: [1.00, 1.25, 512]

# YOLOv8.0 backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]]  # 9

# YOLOv8.0-p2 head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  - [-1, 3, C2f, [512]]  # 12

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  - [-1, 3, C2f, [256]]  # 15 (P3/8-small)

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 2], 1, Concat, [1]]  # cat backbone P2
  - [-1, 3, C2f, [128]]  # 18 (P2/4-xsmall)

  - [-1, 1, Conv, [128, 3, 2]]
  - [[-1, 15], 1, Concat, [1]]  # cat head P3
  - [-1, 3, C2f, [256]]  # 21 (P3/8-small)

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]]  # cat head P4
  - [-1, 3, C2f, [512]]  # 24 (P4/16-medium)

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]]  # cat head P5
  - [-1, 3, C2f, [1024]]  # 27 (P5/32-large)

  - [[18, 21, 24, 27], 1, Detect, [nc]]  # Detect(P2, P3, P4, P5)

yolov8-p2 tensort 实现

参考: https://github.com/wang-xinyu/tensorrtx/tree/master/yolov8

  1. model.cpp 中增加 buildEngineYolov8x_p2 方法.

    Backbone

    • backbone 和 yolov8 一样 , 无需改动，照搬下来就行.

      cpp 复制代码 
          /*******************************************************************************************************
    *****************************************  YOLOV8 BACKBONE  ********************************************
    *******************************************************************************************************/
    nvinfer1::IElementWiseLayer *conv0 = convBnSiLU(network, weightMap, *data, 80, 3, 2, 1, "model.0");
    nvinfer1::IElementWiseLayer *conv1 = convBnSiLU(network, weightMap, *conv0->getOutput(0), 160, 3, 2, 1, "model.1");
    nvinfer1::IElementWiseLayer *conv2 = C2F(network, weightMap, *conv1->getOutput(0), 160, 160, 3, true, 0.5, "model.2");
    nvinfer1::IElementWiseLayer *conv3 = convBnSiLU(network, weightMap, *conv2->getOutput(0), 320, 3, 2, 1, "model.3");
    nvinfer1::IElementWiseLayer *conv4 = C2F(network, weightMap, *conv3->getOutput(0), 320, 320, 6, true, 0.5, "model.4");
    nvinfer1::IElementWiseLayer *conv5 = convBnSiLU(network, weightMap, *conv4->getOutput(0), 640, 3, 2, 1, "model.5");
    nvinfer1::IElementWiseLayer *conv6 = C2F(network, weightMap, *conv5->getOutput(0), 640, 640, 6, true, 0.5, "model.6");
    nvinfer1::IElementWiseLayer *conv7 = convBnSiLU(network, weightMap, *conv6->getOutput(0), 640, 3, 2, 1, "model.7");
    nvinfer1::IElementWiseLayer *conv8 = C2F(network, weightMap, *conv7->getOutput(0), 640, 640, 3, true, 0.5, "model.8");
    nvinfer1::IElementWiseLayer *conv9 = SPPF(network, weightMap, *conv8->getOutput(0), 640, 640, 5, "model.9");

    Head

    • 由3个输出层 (P3、P4、P5) 变成4个输出层 (P2、P3、P4、P5)

      HEAD
      cpp 复制代码 
        /*******************************************************************************************************
    ******************************************  YOLOV8 HEAD  ***********************************************
    *******************************************************************************************************/
    float scale[] = {1.0, 2.0, 2.0};
    nvinfer1::IResizeLayer *upsample10 = network->addResize(*conv9->getOutput(0));
    upsample10->setResizeMode(nvinfer1::ResizeMode::kNEAREST);
    upsample10->setScales(scale, 3);

    nvinfer1::ITensor *inputTensor11[] = {upsample10->getOutput(0), conv6->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat11 = network->addConcatenation(inputTensor11, 2);
    nvinfer1::IElementWiseLayer *conv12 = C2F(network, weightMap, *cat11->getOutput(0), 640, 640, 3, false, 0.5, "model.12");

    nvinfer1::IResizeLayer *upsample13 = network->addResize(*conv12->getOutput(0));
    upsample13->setResizeMode(nvinfer1::ResizeMode::kNEAREST);
    upsample13->setScales(scale, 3);
    nvinfer1::ITensor *inputTensor14[] = {upsample13->getOutput(0), conv4->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat14 = network->addConcatenation(inputTensor14, 2);
    nvinfer1::IElementWiseLayer *conv15 = C2F(network, weightMap, *cat14->getOutput(0), 320, 320, 3, false, 0.5, "model.15");

    nvinfer1::IResizeLayer *upsample16 = network->addResize(*conv15->getOutput(0));
    upsample16->setResizeMode(nvinfer1::ResizeMode::kNEAREST);
    upsample16->setScales(scale, 3);
    nvinfer1::ITensor *inputTensor17[] = {upsample16->getOutput(0), conv2->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat17 = network->addConcatenation(inputTensor17, 2);
    nvinfer1::IElementWiseLayer *conv18 = C2F(network, weightMap, *cat17->getOutput(0), 160, 160, 3, false, 0.5, "model.18");

    nvinfer1::IElementWiseLayer *conv19 = convBnSiLU(network, weightMap, *conv18->getOutput(0), 160, 3, 2, 1, "model.19");
    nvinfer1::ITensor *inputTensor20[] = {conv19->getOutput(0), conv15->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat20 = network->addConcatenation(inputTensor20, 2);
    nvinfer1::IElementWiseLayer *conv21 = C2F(network, weightMap, *cat20->getOutput(0), 320, 320, 3, false, 0.5, "model.21");

    nvinfer1::IElementWiseLayer *conv22 = convBnSiLU(network, weightMap, *conv21->getOutput(0), 320, 3, 2, 1, "model.22");
    nvinfer1::ITensor *inputTensor23[] = {conv22->getOutput(0), conv12->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat23 = network->addConcatenation(inputTensor23, 2);
    nvinfer1::IElementWiseLayer *conv24 = C2F(network, weightMap, *cat23->getOutput(0), 640, 640, 3, false, 0.5, "model.24");

    nvinfer1::IElementWiseLayer *conv25 = convBnSiLU(network, weightMap, *conv24->getOutput(0), 640, 3, 2, 1, "model.25");
    nvinfer1::ITensor *inputTensor26[] = {conv25->getOutput(0), conv9->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat26 = network->addConcatenation(inputTensor26, 2);
    nvinfer1::IElementWiseLayer *conv27 = C2F(network, weightMap, *cat26->getOutput(0), 640, 640, 3, false, 0.5, "model.27");
      OUTPUT
      cpp 复制代码 
      /*******************************************************************************************************
    *********************************************  YOLOV8 OUTPUT  ******************************************
    *******************************************************************************************************/
    // output0
    nvinfer1::IElementWiseLayer *conv28_cv2_0_0 = convBnSiLU(network, weightMap, *conv18->getOutput(0), 64, 3, 1, 1, "model.28.cv2.0.0");
    nvinfer1::IElementWiseLayer *conv28_cv2_0_1 = convBnSiLU(network, weightMap, *conv28_cv2_0_0->getOutput(0), 64, 3, 1, 1, "model.28.cv2.0.1");
    nvinfer1::IConvolutionLayer *conv28_cv2_0_2 = network->addConvolutionNd(*conv28_cv2_0_1->getOutput(0), 64, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv2.0.2.weight"], weightMap["model.28.cv2.0.2.bias"]);
    conv28_cv2_0_2->setStrideNd(nvinfer1::DimsHW{1, 1});
    conv28_cv2_0_2->setPaddingNd(nvinfer1::DimsHW{0, 0});

    nvinfer1::IElementWiseLayer *conv28_cv3_0_0 = convBnSiLU(network, weightMap, *conv18->getOutput(0), 160, 3, 1, 1, "model.28.cv3.0.0");
    nvinfer1::IElementWiseLayer *conv28_cv3_0_1 = convBnSiLU(network, weightMap, *conv28_cv3_0_0->getOutput(0), 160, 3, 1, 1, "model.28.cv3.0.1");
    nvinfer1::IConvolutionLayer *conv28_cv3_0_2 = network->addConvolutionNd(*conv28_cv3_0_1->getOutput(0), kNumClass, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv3.0.2.weight"], weightMap["model.28.cv3.0.2.bias"]);
    conv28_cv3_0_2->setStride(nvinfer1::DimsHW{1, 1});
    conv28_cv3_0_2->setPadding(nvinfer1::DimsHW{0, 0});
    nvinfer1::ITensor *inputTensor28_0[] = {conv28_cv2_0_2->getOutput(0), conv28_cv3_0_2->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_0 = network->addConcatenation(inputTensor28_0, 2); // P2

    // output1
    nvinfer1::IElementWiseLayer *conv28_cv2_1_0 = convBnSiLU(network, weightMap, *conv21->getOutput(0), 64, 3, 1, 1, "model.28.cv2.1.0");
    nvinfer1::IElementWiseLayer *conv28_cv2_1_1 = convBnSiLU(network, weightMap, *conv28_cv2_1_0->getOutput(0), 64, 3, 1, 1, "model.28.cv2.1.1");
    nvinfer1::IConvolutionLayer *conv28_cv2_1_2 = network->addConvolutionNd(*conv28_cv2_1_1->getOutput(0), 64, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv2.1.2.weight"], weightMap["model.28.cv2.1.2.bias"]);
    conv28_cv2_1_2->setStrideNd(nvinfer1::DimsHW{1, 1});
    conv28_cv2_1_2->setPaddingNd(nvinfer1::DimsHW{0, 0});

    nvinfer1::IElementWiseLayer *conv28_cv3_1_0 = convBnSiLU(network, weightMap, *conv21->getOutput(0), 160, 3, 1, 1, "model.28.cv3.1.0");
    nvinfer1::IElementWiseLayer *conv28_cv3_1_1 = convBnSiLU(network, weightMap, *conv28_cv3_1_0->getOutput(0), 160, 3, 1, 1, "model.28.cv3.1.1");
    nvinfer1::IConvolutionLayer *conv28_cv3_1_2 = network->addConvolutionNd(*conv28_cv3_1_1->getOutput(0), kNumClass, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv3.1.2.weight"], weightMap["model.28.cv3.1.2.bias"]);
    conv28_cv3_1_2->setStrideNd(nvinfer1::DimsHW{1, 1});
    conv28_cv3_1_2->setPaddingNd(nvinfer1::DimsHW{0, 0});

    nvinfer1::ITensor *inputTensor28_1[] = {conv28_cv2_1_2->getOutput(0), conv28_cv3_1_2->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_1 = network->addConcatenation(inputTensor28_1, 2);

    // output2
    nvinfer1::IElementWiseLayer *conv28_cv2_2_0 = convBnSiLU(network, weightMap, *conv24->getOutput(0), 64, 3, 1, 1, "model.28.cv2.2.0");
    nvinfer1::IElementWiseLayer *conv28_cv2_2_1 = convBnSiLU(network, weightMap, *conv28_cv2_2_0->getOutput(0), 64, 3, 1, 1, "model.28.cv2.2.1");
    nvinfer1::IConvolutionLayer *conv28_cv2_2_2 = network->addConvolution(*conv28_cv2_2_1->getOutput(0), 64, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv2.2.2.weight"], weightMap["model.28.cv2.2.2.bias"]);

    nvinfer1::IElementWiseLayer *conv28_cv3_2_0 = convBnSiLU(network, weightMap, *conv24->getOutput(0), 160, 3, 1, 1, "model.28.cv3.2.0");
    nvinfer1::IElementWiseLayer *conv28_cv3_2_1 = convBnSiLU(network, weightMap, *conv28_cv3_2_0->getOutput(0), 160, 3, 1, 1, "model.28.cv3.2.1");
    nvinfer1::IConvolutionLayer *conv28_cv3_2_2 = network->addConvolution(*conv28_cv3_2_1->getOutput(0), kNumClass, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv3.2.2.weight"], weightMap["model.28.cv3.2.2.bias"]);

    nvinfer1::ITensor *inputTensor28_2[] = {conv28_cv2_2_2->getOutput(0), conv28_cv3_2_2->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_2 = network->addConcatenation(inputTensor28_2, 2);

    // output3
    nvinfer1::IElementWiseLayer *conv28_cv2_3_0 = convBnSiLU(network, weightMap, *conv27->getOutput(0), 64, 3, 1, 1, "model.28.cv2.3.0");
    nvinfer1::IElementWiseLayer *conv28_cv2_3_1 = convBnSiLU(network, weightMap, *conv28_cv2_3_0->getOutput(0), 64, 3, 1, 1, "model.28.cv2.3.1");
    nvinfer1::IConvolutionLayer *conv28_cv2_3_2 = network->addConvolution(*conv28_cv2_3_1->getOutput(0), 64, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv2.3.2.weight"], weightMap["model.28.cv2.3.2.bias"]);

    nvinfer1::IElementWiseLayer *conv28_cv3_3_0 = convBnSiLU(network, weightMap, *conv27->getOutput(0), 160, 3, 1, 1, "model.28.cv3.3.0");
    nvinfer1::IElementWiseLayer *conv28_cv3_3_1 = convBnSiLU(network, weightMap, *conv28_cv3_3_0->getOutput(0), 160, 3, 1, 1, "model.28.cv3.3.1");
    nvinfer1::IConvolutionLayer *conv28_cv3_3_2 = network->addConvolution(*conv28_cv3_3_1->getOutput(0), kNumClass, nvinfer1::DimsHW{1, 1}, weightMap["model.28.cv3.3.2.weight"], weightMap["model.28.cv3.3.2.bias"]);

    nvinfer1::ITensor *inputTensor28_3[] = {conv28_cv2_3_2->getOutput(0), conv28_cv3_3_2->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_3 = network->addConcatenation(inputTensor28_3, 2);
      DETECT
      cpp 复制代码 
      /*******************************************************************************************************
    *********************************************  YOLOV8 DETECT  ******************************************
    *******************************************************************************************************/
    // P2
    nvinfer1::IShuffleLayer *shuffle28_0 = network->addShuffle(*cat28_0->getOutput(0));
    shuffle28_0->setReshapeDimensions(nvinfer1::Dims2{64 + kNumClass, (kInputH / 4) * (kInputW / 4)});
    nvinfer1::ISliceLayer *split28_0_0 = network->addSlice(*shuffle28_0->getOutput(0), nvinfer1::Dims2{0, 0}, nvinfer1::Dims2{64, (kInputH / 4) * (kInputW / 4)}, nvinfer1::Dims2{1, 1});
    nvinfer1::ISliceLayer *split28_0_1 = network->addSlice(*shuffle28_0->getOutput(0), nvinfer1::Dims2{64, 0}, nvinfer1::Dims2{kNumClass, (kInputH / 4) * (kInputW / 4)}, nvinfer1::Dims2{1, 1});
    nvinfer1::IShuffleLayer *dfl28_0 = DFL(network, weightMap, *split28_0_0->getOutput(0), 4, (kInputH / 4) * (kInputW / 4), 1, 1, 0, "model.28.dfl.conv.weight");
    nvinfer1::ITensor *inputTensor28_dfl_0[] = {dfl28_0->getOutput(0), split28_0_1->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_dfl_0 = network->addConcatenation(inputTensor28_dfl_0, 2);

    // P3
    nvinfer1::IShuffleLayer *shuffle28_1 = network->addShuffle(*cat28_1->getOutput(0));
    shuffle28_1->setReshapeDimensions(nvinfer1::Dims2{64 + kNumClass, (kInputH / 8) * (kInputW / 8)});
    nvinfer1::ISliceLayer *split28_1_0 = network->addSlice(*shuffle28_1->getOutput(0), nvinfer1::Dims2{0, 0}, nvinfer1::Dims2{64, (kInputH / 8) * (kInputW / 8)}, nvinfer1::Dims2{1, 1});
    nvinfer1::ISliceLayer *split28_1_1 = network->addSlice(*shuffle28_1->getOutput(0), nvinfer1::Dims2{64, 0}, nvinfer1::Dims2{kNumClass, (kInputH / 8) * (kInputW / 8)}, nvinfer1::Dims2{1, 1});
    nvinfer1::IShuffleLayer *dfl28_1 = DFL(network, weightMap, *split28_1_0->getOutput(0), 4, (kInputH / 8) * (kInputW / 8), 1, 1, 0, "model.28.dfl.conv.weight");
    nvinfer1::ITensor *inputTensor28_dfl_1[] = {dfl28_1->getOutput(0), split28_1_1->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_dfl_1 = network->addConcatenation(inputTensor28_dfl_1, 2);

    // P4
    nvinfer1::IShuffleLayer *shuffle28_2 = network->addShuffle(*cat28_2->getOutput(0));
    shuffle28_2->setReshapeDimensions(nvinfer1::Dims2{64 + kNumClass, (kInputH / 16) * (kInputW / 16)});
    nvinfer1::ISliceLayer *split28_2_0 = network->addSlice(*shuffle28_2->getOutput(0), nvinfer1::Dims2{0, 0}, nvinfer1::Dims2{64, (kInputH / 16) * (kInputW / 16)}, nvinfer1::Dims2{1, 1});
    nvinfer1::ISliceLayer *split28_2_1 = network->addSlice(*shuffle28_2->getOutput(0), nvinfer1::Dims2{64, 0}, nvinfer1::Dims2{kNumClass, (kInputH / 16) * (kInputW / 16)}, nvinfer1::Dims2{1, 1});
    nvinfer1::IShuffleLayer *dfl28_2 = DFL(network, weightMap, *split28_2_0->getOutput(0), 4, (kInputH / 16) * (kInputW / 16), 1, 1, 0, "model.28.dfl.conv.weight");
    nvinfer1::ITensor *inputTensor28_dfl_2[] = {dfl28_2->getOutput(0), split28_2_1->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_dfl_2 = network->addConcatenation(inputTensor28_dfl_2, 2);

    // P5
    nvinfer1::IShuffleLayer *shuffle28_3 = network->addShuffle(*cat28_3->getOutput(0));
    shuffle28_3->setReshapeDimensions(nvinfer1::Dims2{64 + kNumClass, (kInputH / 32) * (kInputW / 32)});
    nvinfer1::ISliceLayer *split28_3_0 = network->addSlice(*shuffle28_3->getOutput(0), nvinfer1::Dims2{0, 0}, nvinfer1::Dims2{64, (kInputH / 32) * (kInputW / 32)}, nvinfer1::Dims2{1, 1});
    nvinfer1::ISliceLayer *split28_3_1 = network->addSlice(*shuffle28_3->getOutput(0), nvinfer1::Dims2{64, 0}, nvinfer1::Dims2{kNumClass, (kInputH / 32) * (kInputW / 32)}, nvinfer1::Dims2{1, 1});
    nvinfer1::IShuffleLayer *dfl28_3 = DFL(network, weightMap, *split28_3_0->getOutput(0), 4, (kInputH / 32) * (kInputW / 32), 1, 1, 0, "model.28.dfl.conv.weight");
    nvinfer1::ITensor *inputTensor28_dfl_3[] = {dfl28_3->getOutput(0), split28_3_1->getOutput(0)};
    nvinfer1::IConcatenationLayer *cat28_dfl_3 = network->addConcatenation(inputTensor28_dfl_3, 2);

    nvinfer1::IPluginV2Layer *yolo = addYoLoLayer(network, std::vector<nvinfer1::IConcatenationLayer *>{cat28_dfl_0, cat28_dfl_1, cat28_dfl_2, cat28_dfl_3});
    yolo->getOutput(0)->setName(kOutputTensorName);
    network->markOutput(*yolo->getOutput(0));

  2. 修改 yololayer.cu 中 forwardGpu 方法

    cpp 复制代码 
     void YoloLayerPlugin::forwardGpu(const float *const *inputs, float *output, cudaStream_t stream, int mYoloV8netHeight, int mYoloV8NetWidth, int batchSize)
{
   	int outputElem = 1 + mMaxOutObject * sizeof(Detection) / sizeof(float);
    cudaMemsetAsync(output, 0, sizeof(float), stream);
    for (int idx = 0; idx < batchSize; ++idx)
    {
       CUDA_CHECK(cudaMemsetAsync(output + idx * outputElem, 0, sizeof(float), stream));
    }
    int numElem = 0;   
    // int grids[3][2] = {{mYoloV8netHeight / 8, mYoloV8NetWidth / 8}, {mYoloV8netHeight / 16, mYoloV8NetWidth / 16}, {mYoloV8netHeight / 32, mYoloV8NetWidth / 32}};
	// todo 
	int grids[4][2] = {{mYoloV8netHeight / 4, mYoloV8NetWidth / 4}, {mYoloV8netHeight / 8, mYoloV8NetWidth / 8}, {mYoloV8netHeight / 16, mYoloV8NetWidth / 16}, {mYoloV8netHeight / 32, mYoloV8NetWidth / 32}};
	// int strides[] = { 8, 16, 32 };
	// todo 
	int strides[] = {4, 8, 16, 32};
	// for (unsigned int i = 0; i < 3; i++)
	// todo 
    for (unsigned int i = 0; i < 4; i++)
    {
        int grid_h = grids[i][0];
        int grid_w = grids[i][1];
        int stride = strides[i];
        numElem = grid_h * grid_w * batchSize;
        if (numElem < mThreadCount)
            mThreadCount = numElem;

        CalDetection<<<(numElem + mThreadCount - 1) / mThreadCount, mThreadCount, 0, stream>>>(inputs[i], output, numElem, mMaxOutObject, grid_h, grid_w, stride, mClassCount, outputElem);
    }
}

  3. 修改 main.cpp -> serialize_engine ，增加一个 sub_type

    cpp 复制代码 
      ...
  else if (sub_type == "x-p2")
  {
     	serialized_engine = buildEngineYolov8x_p2(builder, config, DataType::kFLOAT, wts_name);
  }
  ...

  4. 参考作者 (https://github.com/wang-xinyu/tensorrtx/tree/master/yolov8) ， 获取wts ， 然后生成模型.
    ./yolov8 -s ./weights/xxx.wts ./weights/xxx.engine x-p2

  5. 推理模型测试
    ./yolov8 -d xxx.engine ../images g

END

  • 官网中没有找到p2的预训练模型，所以需要根据自己数据集训练模型
  • 自己训练模型需要更改 config.h 中对应的参数.
  • 以上纯手工输出，若有不对，欢迎大佬指正.

参考:

相关推荐
一勺汤2 天前
YOLOv8模型改进 第十七讲 通道压缩的自注意力机制CRA
yolo·目标检测·outlook·模块·yolov8·yolov8改进·魔改
就是求关注14 天前
深度学习—基于YOLOv8的人物目标检测和分割（跟踪）
yolo·目标检测·目标跟踪·yolov8·目标分割
你的陈某某19 天前
Atlas800昇腾服务器（型号：3000）—YOLO全系列NPU推理【跟踪】（八）
yolov8·npu·bytetrack·atlas800·a300i pro·ais_bench
aworkholic21 天前
opencv dnn模块 示例(27) 目标检测 object_detection 之 yolov11
pytorch·opencv·yolo·目标检测·dnn·tensorrt·yolo11
machnerrn24 天前
基于YOLOv9实现的自行车检测系统：为共享自行车违停项目开发（附项目源码及数据集下载）
深度学习·目标检测·毕业设计·课程设计·yolov8·yolov9·自行车违停检测
知来者逆1 个月前
基于YOLOv8目标检测与chef-transformer（T5）从图像创建食谱
人工智能·深度学习·yolo·目标检测·计算机视觉·transformer·yolov8
cheoyeon1 个月前
国产化框架PaddleYOLO结合Swanlab进行作物检测
深度学习·目标检测·作物检测·国产化·yolov8·swanlab·粮食安全
刘悦的技术博客1 个月前
Win11本地部署FaceFusion3最强AI换脸,集成Tensorrt10.4推理加速,让甜品显卡也能发挥生产力
ai·tensorrt·cuda
墨理学AI1 个月前
Mac 电脑配置yolov8运行环境实现目标追踪、计数、画出轨迹、多线程
yolo·macos·yolov8
不想敲代码！！！2 个月前
爆改YOLOv8|使用MobileNetV4替换yolov8的Backbone
pytorch·python·深度学习·yolo·目标检测·yolov8
热门推荐
01【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总02(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明03组基轨迹建模 GBTM的介绍与实现（Stata 或 R）04【AIGC】重塑未来的科技巨轮05全面解析：构建基于深度学习的安全帽检测系统（UI界面+YOLO代码+数据集）06【经验分享】Ubuntu22.04安装微信(linux官方版)07基于YOLOv10深度学习的CT扫描图像肾结石智能检测系统【python源码+Pyqt5界面+数据集+训练代码】深度学习实战、目标检测08Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO09RAG 实践- Ollama+RagFlow 部署本地知识库10红米手机使用google play