5. 通用yolov7模型部署¶

5.1. 引言¶

本文档介绍了如何将yolov7架构的模型部署在cv181x开发板的操作流程，主要的操作步骤包括：

yolov7模型pytorch版本转换为onnx模型
onnx模型转换为cvimodel格式
最后编写调用接口获取推理结果

5.2. pt模型转换为onnx¶

下载官方[yolov7](https://github.com/WongKinYiu/yolov7)仓库代码

git clone https://github.com/WongKinYiu/yolov7.git

在上述下载代码的目录中新建一个文件夹weights，然后将需要导出onnx的模型移动到yolov7/weights

cd yolov7 & mkdir weights
cp path/to/onnx ./weights/

然后将yolo_export/yolov7_export.py复制到yolov7目录下

然后使用以下命令导出TDL_SDK形式的yolov7模型

python yolov7_export.py --weights ./weights/yolov7-tiny.pt

小技巧

如果输入为1080p的视频流，建议将模型输入尺寸改为384x640，可以减少冗余计算，提高推理速度，如下命令所示：

python yolov7_export.py --weights ./weights/yolov7-tiny.pt --img-size 384 640

yolo_export中的脚本可以通过SFTP获取：下载站台:sftp://218.17.249.213 帐号:cvitek_mlir_2023 密码:7&2Wd%cu5k

通过SFTP找到下图对应的文件夹：

5.3. onnx模型转换cvimodel¶

cvimodel转换操作可以参考yolo-v5移植章节的onnx模型转换cvimodel部分。

小心

yolov7官方版本的模型预处理参数，即mean以及scale与yolov5相同，可以复用yolov5转换cvimodel的命令

5.4. TDL_SDK接口说明¶

yolov7模型与yolov5模型检测与解码过程基本类似，主要不同是anchor的不同

小心

注意修改anchors为yolov7的anchors!!!

anchors:

[12,16, 19,36, 40,28] # P3/8
[36,75, 76,55, 72,146] # P4/16
[142,110, 192,243, 459,401] # P5/32

预处理接口设置如下代码所示

// set preprocess and algorithm param for yolov7 detection
// if use official model, no need to change param
CVI_S32 init_param(const cvitdl_handle_t tdl_handle) {
  // setup preprocess
  YoloPreParam preprocess_cfg =
      CVI_TDL_Get_YOLO_Preparam(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7);

  for (int i = 0; i < 3; i++) {
    printf("asign val %d \n", i);
    preprocess_cfg.factor[i] = 0.003922;
    preprocess_cfg.mean[i] = 0.0;
  }
  preprocess_cfg.format = PIXEL_FORMAT_RGB_888_PLANAR;

  printf("setup yolov7 param \n");
  CVI_S32 ret =
      CVI_TDL_Set_YOLO_Preparam(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7, preprocess_cfg);
  if (ret != CVI_SUCCESS) {
    printf("Can not set Yolov5 preprocess parameters %#x\n", ret);
    return ret;
  }

  // setup yolo algorithm preprocess
  YoloAlgParam yolov7_param = CVI_TDL_Get_YOLO_Algparam(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7);
  uint32_t *anchors = new uint32_t[18];
  uint32_t p_anchors[18] = {12, 16, 19,  36,  40,  28,  36,  75,  76,
                            55, 72, 146, 142, 110, 192, 243, 459, 401};
  memcpy(anchors, p_anchors, sizeof(p_anchors));
  yolov7_param.anchors = anchors;

  uint32_t *strides = new uint32_t[3];
  uint32_t p_strides[3] = {8, 16, 32};
  memcpy(strides, p_strides, sizeof(p_strides));
  yolov7_param.strides = strides;
  yolov7_param.cls = 80;

  printf("setup yolov7 algorithm param \n");
  ret = CVI_TDL_Set_YOLO_Algparam(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7, yolov7_param);
  if (ret != CVI_SUCCESS) {
    printf("Can not set Yolov5 algorithm parameters %#x\n", ret);
    return ret;
  }

  // set thershold
  CVI_TDL_SetModelThreshold(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7, 0.5);
  CVI_TDL_SetModelNmsThreshold(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7, 0.5);

  printf("yolov7 algorithm parameters setup success!\n");
  return ret;
}

推理接口如下所示：

ret = CVI_TDL_OpenModel(tdl_handle, CVI_TDL_SUPPORTED_MODEL_YOLOV7, model_path.c_str());
if (ret != CVI_SUCCESS) {
  printf("open model failed %#x!\n", ret);
  return ret;
}

std::cout << "model opened:" << model_path << std::endl;

VIDEO_FRAME_INFO_S fdFrame;
ret = CVI_TDL_ReadImage(str_src_dir.c_str(), &fdFrame, PIXEL_FORMAT_RGB_888);
std::cout << "CVI_TDL_ReadImage done!\n";

if (ret != CVI_SUCCESS) {
  std::cout << "Convert out video frame failed with :" << ret << ".file:" << str_src_dir
            << std::endl;
  // continue;
}

cvtdl_object_t obj_meta = {0};

CVI_TDL_Yolov7(tdl_handle, &fdFrame, &obj_meta);

for (uint32_t i = 0; i < obj_meta.size; i++) {
  printf("detect res: %f %f %f %f %f %d\n", obj_meta.info[i].bbox.x1, obj_meta.info[i].bbox.y1,
         obj_meta.info[i].bbox.x2, obj_meta.info[i].bbox.y2, obj_meta.info[i].bbox.score,
         obj_meta.info[i].classes);
}

5.5. 测试结果¶

测试了yolov7-tiny模型各个版本的指标，测试数据为COCO2017，其中阈值设置为：

conf_threshold: 0.001
nms_threshold: 0.65

分辨率均为640 x 640

yolov7-tiny模型的官方导出方式性能：

测试平台	推理耗时 (ms)	带宽 (MB)	ION(MB)	MAP 0.5	MAP 0.5-0.95
pytorch	N/A	N/A	N/A	56.7	38.7
cv180x	ion分配失败	ion分配失败	38.97	量化失败	量化失败
cv181x	75.4	85.31	17.54	量化失败	量化失败
cv182x	56.6	85.31	17.54	量化失败	量化失败
cv183x	21.85	71.46	16.15	量化失败	量化失败
cv186x	7.91	137.72	23.87	量化失败	量化失败

yolov7-tiny模型的TDL_SDK导出方式性能：

测试平台	推理耗时 (ms)	带宽 (MB)	ION(MB)	MAP 0.5	MAP 0.5-0.95
onnx	N/A	N/A	N/A	53.7094	36.438
cv180x	ion分配失败	ion分配失败	36.81	ion分配失败	ion分配失败
cv181x	70.41	70.66	15.43	53.3681	32.6277
cv182x	52.01	70.66	15.43	53.3681	32.6277
cv183x	18.95	55.86	14.05	53.3681	32.6277
cv186x	6.54	99.41	17.98	53.44	33.08