RetinaFace MXNet模型转ONNX转TensorRT

文章目录

RetinaFace MXNet模型转ONNX转TensorRT

1. github开源代码

RetinaFace TensorRT推理的开源代码位置在https://github.com/linghu8812/tensorrt_inference/tree/master/RetinaFace

2. MXNet模型转ONNX模型

首先通过命令git clone https://github.com/deepinsight/insightface.gitclone insightface的代码,然后将export_onnx.py文件拷贝到./detection/RetinaFace或者./detection/RetinaFaceAntiCov文件夹中,通过以下命令生成ONNX文件。对于RetinaFace-R50RetinaFace-MobileNet0.25RetinaFaceAntiCov这几个模型都可以支持。通过以下命令可以导出模型:

  • 导出resnet50模型
python3 export_onnx.py
  • 导出mobilenet 0.25模型
python3 export_onnx.py  --prefix ./model/mnet.25
  • 导出RetinaFaceAntiCov模型
python3 export_onnx.py  --prefix ./model/mnet_cov2 --network net3l

YOLOv4模型一样,对输出结果也做了concat,如下图所示。
RetinaFace MXNet模型转ONNX转TensorRT

3. ONNX模型转TensorRT模型

3.1 概述

TensorRT模型即TensorRT的推理引擎,代码中通过C++实现。相关配置写在config.yaml文件中,如果存在engine_file的路径,则读取engine_file,否则从onnx_file生成engine_file

void RetinaFace::LoadEngine() {
    // create and load engine
    std::fstream existEngine;
    existEngine.open(engine_file, std::ios::in);
    if (existEngine) {
        readTrtFile(engine_file, engine);
        assert(engine != nullptr);
    } else {
        onnxToTRTModel(onnx_file, engine_file, engine, BATCH_SIZE);
        assert(engine != nullptr);
    }
}

config.yaml文件可以设置batch size,图像的size及模型的anchor等。

RetinaFace:
    onnx_file:     "../R50.onnx"
    engine_file:   "../R50.trt"
    BATCH_SIZE:    1
    INPUT_CHANNEL: 3
    IMAGE_WIDTH:   640
    IMAGE_HEIGHT:  640
    obj_threshold: 0.5
    nms_threshold: 0.45
    detect_mask:   False
    mask_thresh:   0.5
    landmark_std:  1
    feature_steps: [32, 16, 8]
    anchor_sizes:  [[512, 256], [128, 64], [32, 16]]

3.2 编译

通过以下命令对项目进行编译,生成RetinaFace_trt

mkdir build && cd build
cmake ..
make -j

3.3 运行

通过以下命令运行项目,得到推理结果

  • RetinaFace模型推理
./RetinaFace_trt../config.yaml ../samples
  • RetinaFaceAntiCov模型推理
./RetinaFace_trt ../config_anti.yaml ../samples

4. 推理结果

  • RetinaFace推理结果:
    RetinaFace MXNet模型转ONNX转TensorRT
  • RetinaFaceAntiCov推理结果:
    RetinaFace MXNet模型转ONNX转TensorRT
上一篇:pytorch模型结构可视化,可显示每层的尺寸


下一篇:onnx算子大全