在ubuntu24.04运行yolov5（用CPU）

hell没有哦2025-12-11 18:31

在ubuntu24.04运行yolov5（用CPU）修正1

环境
开发机：Ubuntu 24.04（刷系统，用真实主机跑ubuntu，效率很快）

  • 目标板：ubuntu24
  • 已有：yolov5n.pt
  • 自己写的文档，有点长，如果能在ubuntu跑通，放在其它板子换个交叉编译即可（有用点点赞哦）

一、思路架构

  • Windows（GPU） ：只做 pt -> torchscript（导出速度更快、PyTorch 环境你也说更成熟）
  • Ubuntu（CPU） ：做 torchscript -> ncnn(param/bin) + 编译/运行

1、window10操作

1.1、使用window10将PyTorch导出TorchScript

1.1.1、创建脚本export_ts.py

有一个前置条件：在window10部署能够跑yolo的环境，才能用下列脚本，不然应该会报错

复制代码 
import torch

print("loading yolov5n...")
model = torch.hub.load('ultralytics/yolov5', 'yolov5n', pretrained=True)
model.eval()

dummy = torch.randn(1, 3, 640, 640)

print("tracing...")
ts = torch.jit.trace(model, dummy)
ts.save("yolov5n.torchscript.pt")

print("saved --> yolov5n.torchscript.pt")

1.1.2、运行脚本

复制代码 
python export_ts.py

1.1.3、脚本结果

复制代码 
yolov5n.torchscript.pt

1.2、在ubuntu检测结果：是正常的 TorchScript（如下）

复制代码 
user@user-EliteMini-Series:~$ file yolov5n.torchscript 
yolov5n.torchscript: Zip archive data, at least v0.0 to extract, compression method=store

1.3、问题：为什么生成是 yolov5s.torchscript 而不是 yolov5s.torchscript.pt

复制代码 
1. YOLOv5的export.py在不同版本、不同参数组合下，TorchScript 产物文件名可能是：yolov5s.torchscript、yolov5s.torchscript.pt、yolov5s.torchscript.zip
2. TorchScript 本质通常是一个 zip 打包的模型（TorchScript archive），后缀不影响内容。
3. pnnx 接收的是 TorchScript 模型文件路径，直接把yolov5s.torchscript作为输入也可以。

2、ubuntu操作

2.1、ubuntu依赖环境

bash 复制代码 
sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libomp-dev libvulkan-dev vulkan-tools libopencv-dev python3 python3-pip

2.2、ubuntu下载源码

复制代码 
cd /home/user/RK/RK3568/4_rk3568_ncnn
mkdir -p rk3568_work
cd rk3568_work

git clone https://github.com/Tencent/ncnn.git
cd ncnn
git submodule update --init

2.3、安装虚拟环境将TorchScript使用PNNX转NCNN

复制代码 
# 创建一个新的虚拟环境
python3 -m venv /home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/my_pytorch_env
# 激活虚拟环境
source /home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/my_pytorch_env/bin/activate
pip install torch torchvision torchaudio
pip install pnnx
pnnx yolov5n.torchscript inputshape=[1,3,640,640] fp16=0 optlevel=2 device=cpu

转换后的模型

bash 复制代码 
(my_pytorch_env) user@user-EliteMini-Series:~/RK/RK3568/4_rk3568_ncnn/rk3568_work$ ls
app-rk3568      ncnn-build-host    yolov5n.ncnn.param  yolov5n.pnnx.onnx   yolov5n.torchscript
my_pytorch_env  ncnn-build-rk3568  yolov5n_ncnn.py     yolov5n.pnnx.param
ncnn            yolov5n.ncnn.bin   yolov5n.pnnx.bin    yolov5n_pnnx.py

转换后的模型功能

复制代码 
yolov5n.torchscript        # 你的 TorchScript 原始模型
yolov5n.ncnn.param         # ✅ 最终 ncnn 模型（param）--最重要
yolov5n.ncnn.bin           # ✅ 最终 ncnn 模型（bin）--最重要
yolov5n.pnnx.onnx          # 中间 onnx
yolov5n.pnnx.param         # 中间 pnnx 版 ncnn（一般是更原始图）
yolov5n.pnnx.bin
yolov5n_pnnx.py            # pnnx 产生的 python demo（吃 pnnx.param/bin）
yolov5n_ncnn.py            # pnnx 产生的 python demo（吃 ncnn.param/bin）

C++程序就这么写

复制代码 
net.load_param("yolov5n.ncnn.param");
net.load_model("yolov5n.ncnn.bin");

测试PNNX转换NCNN（如下测试，可选）
compare.py

复制代码 
import torch
from yolov5n_ncnn import test_inference as ncnn_infer  # 导入 NCNN 推理函数
from yolov5n_pnnx import Model as PNNXModel           # 导入 PNNX 模型
from yolov5n_pnnx import test_inference  # 从 yolov5n_pnnx 导入 test_inference 函数

def compare_outputs():
    # 为了保证一致，统一使用一个输入
    torch.manual_seed(0)
    inp = torch.rand(1, 3, 640, 640, dtype=torch.float)

    # 1. PNNX 推理
    pnnx_model = PNNXModel()
    pnnx_output = test_inference(pnnx_model, inp)  # 使用相同输入

    # 如果 pnnx_output 是元组，取出第一个元素
    if isinstance(pnnx_output, tuple):
        pnnx_output = pnnx_output[0]  # 取元组中的第一个元素

    # 2. NCNN 推理
    ncnn_output = ncnn_infer(inp)  # 传入相同输入

    # 3. 检查维度是否一致
    print("pnnx shape:", pnnx_output.shape)
    print("ncnn shape:", ncnn_output.shape)
    print("Shapes equal:", pnnx_output.shape == ncnn_output.shape)

    # 4. 计算最大绝对误差
    diff = (pnnx_output - ncnn_output).abs()
    print("Max absolute diff:", diff.max().item())

    # 5. 使用 allclose 检查是否对齐
    allclose = torch.allclose(pnnx_output, ncnn_output, rtol=1e-3, atol=1e-6)
    print("Allclose (within 1e-6):", allclose)

if __name__ == "__main__":
    compare_outputs()

yolov5n_ncnn.py

yolov5n_ncnn.py 复制代码 
# yolov5n_ncnn.py
import torch
import ncnn
import numpy as np

def test_inference(in0=None):
    if in0 is None:
        torch.manual_seed(0)
        in0 = torch.rand(1, 3, 640, 640, dtype=torch.float)

    out = []

    with ncnn.Net() as net:
        net.load_param("yolov5n.ncnn.param")
        net.load_model("yolov5n.ncnn.bin")

        with net.create_extractor() as ex:
            # NCNN 需要在输入时转换数据类型
            ex.input("in0", ncnn.Mat(in0.squeeze(0).numpy()).clone())

            _, out0 = ex.extract("out0")
            out.append(torch.from_numpy(np.array(out0)).unsqueeze(0))

    if len(out) == 1:
        return out[0]
    else:
        return tuple(out)

if __name__ == "__main__":
    print(test_inference())

yolov5n_pnnx.py

yolov5n_pnnx.py 复制代码 
import os
import numpy as np
import tempfile, zipfile
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
    import torchvision
    import torchaudio
except:
    pass

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

        self.model_0_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=3, kernel_size=(6,6), out_channels=16, padding=(2,2), padding_mode='zeros', stride=(2,2))
        self.model_0_act = nn.SiLU()
        self.model_1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=16, kernel_size=(3,3), out_channels=32, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_1_act = nn.SiLU()
        self.model_2_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=16, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_2_cv1_act = nn.SiLU()
        self.model_2_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=16, kernel_size=(1,1), out_channels=16, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_2_m_0_cv1_act = nn.SiLU()
        self.model_2_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=16, kernel_size=(3,3), out_channels=16, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_2_m_0_cv2_act = nn.SiLU()
        self.model_2_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=16, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_2_cv2_act = nn.SiLU()
        self.model_2_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_2_cv3_act = nn.SiLU()
        self.model_3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_3_act = nn.SiLU()
        self.model_4_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_4_cv1_act = nn.SiLU()
        self.model_4_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_4_m_0_cv1_act = nn.SiLU()
        self.model_4_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(3,3), out_channels=32, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_4_m_0_cv2_act = nn.SiLU()
        self.model_4_m_1_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_4_m_1_cv1_act = nn.SiLU()
        self.model_4_m_1_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(3,3), out_channels=32, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_4_m_1_cv2_act = nn.SiLU()
        self.model_4_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_4_cv2_act = nn.SiLU()
        self.model_4_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_4_cv3_act = nn.SiLU()
        self.model_5_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=128, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_5_act = nn.SiLU()
        self.model_6_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_cv1_act = nn.SiLU()
        self.model_6_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_m_0_cv1_act = nn.SiLU()
        self.model_6_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_6_m_0_cv2_act = nn.SiLU()
        self.model_6_m_1_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_m_1_cv1_act = nn.SiLU()
        self.model_6_m_1_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_6_m_1_cv2_act = nn.SiLU()
        self.model_6_m_2_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_m_2_cv1_act = nn.SiLU()
        self.model_6_m_2_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_6_m_2_cv2_act = nn.SiLU()
        self.model_6_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_cv2_act = nn.SiLU()
        self.model_6_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_6_cv3_act = nn.SiLU()
        self.model_7_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(3,3), out_channels=256, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_7_act = nn.SiLU()
        self.model_8_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_8_cv1_act = nn.SiLU()
        self.model_8_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_8_m_0_cv1_act = nn.SiLU()
        self.model_8_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(3,3), out_channels=128, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_8_m_0_cv2_act = nn.SiLU()
        self.model_8_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_8_cv2_act = nn.SiLU()
        self.model_8_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=256, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_8_cv3_act = nn.SiLU()
        self.model_9_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_9_cv1_act = nn.SiLU()
        self.model_9_m = nn.MaxPool2d(ceil_mode=False, dilation=(1,1), kernel_size=(5,5), padding=(2,2), return_indices=False, stride=(1,1))
        self.pnnx_unique_0 = nn.MaxPool2d(ceil_mode=False, dilation=(1,1), kernel_size=(5,5), padding=(2,2), return_indices=False, stride=(1,1))
        self.pnnx_unique_1 = nn.MaxPool2d(ceil_mode=False, dilation=(1,1), kernel_size=(5,5), padding=(2,2), return_indices=False, stride=(1,1))
        self.model_9_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=512, kernel_size=(1,1), out_channels=256, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_9_cv2_act = nn.SiLU()
        self.model_10_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_10_act = nn.SiLU()
        self.model_11 = nn.Upsample(mode='nearest', scale_factor=(2.0,2.0), size=None)
        self.model_13_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_13_cv1_act = nn.SiLU()
        self.model_13_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_13_m_0_cv1_act = nn.SiLU()
        self.model_13_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_13_m_0_cv2_act = nn.SiLU()
        self.model_13_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_13_cv2_act = nn.SiLU()
        self.model_13_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_13_cv3_act = nn.SiLU()
        self.model_14_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_14_act = nn.SiLU()
        self.model_15 = nn.Upsample(mode='nearest', scale_factor=(2.0,2.0), size=None)
        self.model_17_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_17_cv1_act = nn.SiLU()
        self.model_17_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_17_m_0_cv1_act = nn.SiLU()
        self.model_17_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=32, kernel_size=(3,3), out_channels=32, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_17_m_0_cv2_act = nn.SiLU()
        self.model_17_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=32, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_17_cv2_act = nn.SiLU()
        self.model_17_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_17_cv3_act = nn.SiLU()
        self.model_18_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_18_act = nn.SiLU()
        self.model_20_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_20_cv1_act = nn.SiLU()
        self.model_20_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_20_m_0_cv1_act = nn.SiLU()
        self.model_20_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(3,3), out_channels=64, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_20_m_0_cv2_act = nn.SiLU()
        self.model_20_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=64, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_20_cv2_act = nn.SiLU()
        self.model_20_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_20_cv3_act = nn.SiLU()
        self.model_21_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(3,3), out_channels=128, padding=(1,1), padding_mode='zeros', stride=(2,2))
        self.model_21_act = nn.SiLU()
        self.model_23_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_23_cv1_act = nn.SiLU()
        self.model_23_m_0_cv1_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_23_m_0_cv1_act = nn.SiLU()
        self.model_23_m_0_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(3,3), out_channels=128, padding=(1,1), padding_mode='zeros', stride=(1,1))
        self.model_23_m_0_cv2_act = nn.SiLU()
        self.model_23_cv2_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=128, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_23_cv2_act = nn.SiLU()
        self.model_23_cv3_conv = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=256, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_23_cv3_act = nn.SiLU()
        self.model_24_m_0 = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=64, kernel_size=(1,1), out_channels=255, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_24_m_1 = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=128, kernel_size=(1,1), out_channels=255, padding=(0,0), padding_mode='zeros', stride=(1,1))
        self.model_24_m_2 = nn.Conv2d(bias=True, dilation=(1,1), groups=1, in_channels=256, kernel_size=(1,1), out_channels=255, padding=(0,0), padding_mode='zeros', stride=(1,1))

        archive = zipfile.ZipFile('yolov5n.pnnx.bin', 'r')
        self.model_0_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.0.conv.bias', (16), 'float32')
        self.model_0_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.0.conv.weight', (16,3,6,6), 'float32')
        self.model_1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.1.conv.bias', (32), 'float32')
        self.model_1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.1.conv.weight', (32,16,3,3), 'float32')
        self.model_2_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv1.conv.bias', (16), 'float32')
        self.model_2_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv1.conv.weight', (16,32,1,1), 'float32')
        self.model_2_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.2.m.0.cv1.conv.bias', (16), 'float32')
        self.model_2_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.2.m.0.cv1.conv.weight', (16,16,1,1), 'float32')
        self.model_2_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.2.m.0.cv2.conv.bias', (16), 'float32')
        self.model_2_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.2.m.0.cv2.conv.weight', (16,16,3,3), 'float32')
        self.model_2_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv2.conv.bias', (16), 'float32')
        self.model_2_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv2.conv.weight', (16,32,1,1), 'float32')
        self.model_2_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv3.conv.bias', (32), 'float32')
        self.model_2_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.2.cv3.conv.weight', (32,32,1,1), 'float32')
        self.model_3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.3.conv.bias', (64), 'float32')
        self.model_3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.3.conv.weight', (64,32,3,3), 'float32')
        self.model_4_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv1.conv.bias', (32), 'float32')
        self.model_4_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv1.conv.weight', (32,64,1,1), 'float32')
        self.model_4_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.0.cv1.conv.bias', (32), 'float32')
        self.model_4_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.0.cv1.conv.weight', (32,32,1,1), 'float32')
        self.model_4_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.0.cv2.conv.bias', (32), 'float32')
        self.model_4_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.0.cv2.conv.weight', (32,32,3,3), 'float32')
        self.model_4_m_1_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.1.cv1.conv.bias', (32), 'float32')
        self.model_4_m_1_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.1.cv1.conv.weight', (32,32,1,1), 'float32')
        self.model_4_m_1_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.1.cv2.conv.bias', (32), 'float32')
        self.model_4_m_1_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.m.1.cv2.conv.weight', (32,32,3,3), 'float32')
        self.model_4_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv2.conv.bias', (32), 'float32')
        self.model_4_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv2.conv.weight', (32,64,1,1), 'float32')
        self.model_4_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv3.conv.bias', (64), 'float32')
        self.model_4_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.4.cv3.conv.weight', (64,64,1,1), 'float32')
        self.model_5_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.5.conv.bias', (128), 'float32')
        self.model_5_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.5.conv.weight', (128,64,3,3), 'float32')
        self.model_6_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv1.conv.bias', (64), 'float32')
        self.model_6_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv1.conv.weight', (64,128,1,1), 'float32')
        self.model_6_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.0.cv1.conv.bias', (64), 'float32')
        self.model_6_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.0.cv1.conv.weight', (64,64,1,1), 'float32')
        self.model_6_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.0.cv2.conv.bias', (64), 'float32')
        self.model_6_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.0.cv2.conv.weight', (64,64,3,3), 'float32')
        self.model_6_m_1_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.1.cv1.conv.bias', (64), 'float32')
        self.model_6_m_1_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.1.cv1.conv.weight', (64,64,1,1), 'float32')
        self.model_6_m_1_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.1.cv2.conv.bias', (64), 'float32')
        self.model_6_m_1_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.1.cv2.conv.weight', (64,64,3,3), 'float32')
        self.model_6_m_2_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.2.cv1.conv.bias', (64), 'float32')
        self.model_6_m_2_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.2.cv1.conv.weight', (64,64,1,1), 'float32')
        self.model_6_m_2_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.2.cv2.conv.bias', (64), 'float32')
        self.model_6_m_2_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.m.2.cv2.conv.weight', (64,64,3,3), 'float32')
        self.model_6_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv2.conv.bias', (64), 'float32')
        self.model_6_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv2.conv.weight', (64,128,1,1), 'float32')
        self.model_6_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv3.conv.bias', (128), 'float32')
        self.model_6_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.6.cv3.conv.weight', (128,128,1,1), 'float32')
        self.model_7_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.7.conv.bias', (256), 'float32')
        self.model_7_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.7.conv.weight', (256,128,3,3), 'float32')
        self.model_8_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv1.conv.bias', (128), 'float32')
        self.model_8_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv1.conv.weight', (128,256,1,1), 'float32')
        self.model_8_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.8.m.0.cv1.conv.bias', (128), 'float32')
        self.model_8_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.8.m.0.cv1.conv.weight', (128,128,1,1), 'float32')
        self.model_8_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.8.m.0.cv2.conv.bias', (128), 'float32')
        self.model_8_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.8.m.0.cv2.conv.weight', (128,128,3,3), 'float32')
        self.model_8_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv2.conv.bias', (128), 'float32')
        self.model_8_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv2.conv.weight', (128,256,1,1), 'float32')
        self.model_8_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv3.conv.bias', (256), 'float32')
        self.model_8_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.8.cv3.conv.weight', (256,256,1,1), 'float32')
        self.model_9_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.9.cv1.conv.bias', (128), 'float32')
        self.model_9_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.9.cv1.conv.weight', (128,256,1,1), 'float32')
        self.model_9_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.9.cv2.conv.bias', (256), 'float32')
        self.model_9_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.9.cv2.conv.weight', (256,512,1,1), 'float32')
        self.model_10_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.10.conv.bias', (128), 'float32')
        self.model_10_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.10.conv.weight', (128,256,1,1), 'float32')
        self.model_13_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv1.conv.bias', (64), 'float32')
        self.model_13_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv1.conv.weight', (64,256,1,1), 'float32')
        self.model_13_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.13.m.0.cv1.conv.bias', (64), 'float32')
        self.model_13_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.13.m.0.cv1.conv.weight', (64,64,1,1), 'float32')
        self.model_13_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.13.m.0.cv2.conv.bias', (64), 'float32')
        self.model_13_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.13.m.0.cv2.conv.weight', (64,64,3,3), 'float32')
        self.model_13_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv2.conv.bias', (64), 'float32')
        self.model_13_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv2.conv.weight', (64,256,1,1), 'float32')
        self.model_13_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv3.conv.bias', (128), 'float32')
        self.model_13_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.13.cv3.conv.weight', (128,128,1,1), 'float32')
        self.model_14_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.14.conv.bias', (64), 'float32')
        self.model_14_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.14.conv.weight', (64,128,1,1), 'float32')
        self.model_17_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv1.conv.bias', (32), 'float32')
        self.model_17_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv1.conv.weight', (32,128,1,1), 'float32')
        self.model_17_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.17.m.0.cv1.conv.bias', (32), 'float32')
        self.model_17_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.17.m.0.cv1.conv.weight', (32,32,1,1), 'float32')
        self.model_17_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.17.m.0.cv2.conv.bias', (32), 'float32')
        self.model_17_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.17.m.0.cv2.conv.weight', (32,32,3,3), 'float32')
        self.model_17_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv2.conv.bias', (32), 'float32')
        self.model_17_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv2.conv.weight', (32,128,1,1), 'float32')
        self.model_17_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv3.conv.bias', (64), 'float32')
        self.model_17_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.17.cv3.conv.weight', (64,64,1,1), 'float32')
        self.model_18_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.18.conv.bias', (64), 'float32')
        self.model_18_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.18.conv.weight', (64,64,3,3), 'float32')
        self.model_20_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv1.conv.bias', (64), 'float32')
        self.model_20_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv1.conv.weight', (64,128,1,1), 'float32')
        self.model_20_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.20.m.0.cv1.conv.bias', (64), 'float32')
        self.model_20_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.20.m.0.cv1.conv.weight', (64,64,1,1), 'float32')
        self.model_20_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.20.m.0.cv2.conv.bias', (64), 'float32')
        self.model_20_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.20.m.0.cv2.conv.weight', (64,64,3,3), 'float32')
        self.model_20_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv2.conv.bias', (64), 'float32')
        self.model_20_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv2.conv.weight', (64,128,1,1), 'float32')
        self.model_20_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv3.conv.bias', (128), 'float32')
        self.model_20_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.20.cv3.conv.weight', (128,128,1,1), 'float32')
        self.model_21_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.21.conv.bias', (128), 'float32')
        self.model_21_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.21.conv.weight', (128,128,3,3), 'float32')
        self.model_23_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv1.conv.bias', (128), 'float32')
        self.model_23_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv1.conv.weight', (128,256,1,1), 'float32')
        self.model_23_m_0_cv1_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.23.m.0.cv1.conv.bias', (128), 'float32')
        self.model_23_m_0_cv1_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.23.m.0.cv1.conv.weight', (128,128,1,1), 'float32')
        self.model_23_m_0_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.23.m.0.cv2.conv.bias', (128), 'float32')
        self.model_23_m_0_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.23.m.0.cv2.conv.weight', (128,128,3,3), 'float32')
        self.model_23_cv2_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv2.conv.bias', (128), 'float32')
        self.model_23_cv2_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv2.conv.weight', (128,256,1,1), 'float32')
        self.model_23_cv3_conv.bias = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv3.conv.bias', (256), 'float32')
        self.model_23_cv3_conv.weight = self.load_pnnx_bin_as_parameter(archive, 'model.23.cv3.conv.weight', (256,256,1,1), 'float32')
        self.model_24_m_0.bias = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.0.bias', (255), 'float32')
        self.model_24_m_0.weight = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.0.weight', (255,64,1,1), 'float32')
        self.model_24_m_1.bias = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.1.bias', (255), 'float32')
        self.model_24_m_1.weight = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.1.weight', (255,128,1,1), 'float32')
        self.model_24_m_2.bias = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.2.bias', (255), 'float32')
        self.model_24_m_2.weight = self.load_pnnx_bin_as_parameter(archive, 'model.24.m.2.weight', (255,256,1,1), 'float32')
        self.pnnx_53_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_53.data', (1,3,20,20,2,), 'float32')
        self.pnnx_54_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_54.data', (1,3,20,20,2,), 'float32')
        self.pnnx_55_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_55.data', (1,3,40,40,2,), 'float32')
        self.pnnx_56_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_56.data', (1,3,40,40,2,), 'float32')
        self.pnnx_58_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_58.data', (1,3,80,80,2,), 'float32')
        self.pnnx_60_data = self.load_pnnx_bin_as_parameter(archive, 'pnnx_60.data', (1,3,80,80,2,), 'float32')
        archive.close()

    def load_pnnx_bin_as_parameter(self, archive, key, shape, dtype, requires_grad=True):
        return nn.Parameter(self.load_pnnx_bin_as_tensor(archive, key, shape, dtype), requires_grad)

    def load_pnnx_bin_as_tensor(self, archive, key, shape, dtype):
        fd, tmppath = tempfile.mkstemp()
        with os.fdopen(fd, 'wb') as tmpf, archive.open(key) as keyfile:
            tmpf.write(keyfile.read())
        m = np.memmap(tmppath, dtype=dtype, mode='r', shape=shape).copy()
        os.remove(tmppath)
        return torch.from_numpy(m)

    def forward(self, v_0):
        v_1 = self.model_0_conv(v_0)
        v_2 = self.model_0_act(v_1)
        v_3 = self.model_1_conv(v_2)
        v_4 = self.model_1_act(v_3)
        v_5 = self.model_2_cv1_conv(v_4)
        v_6 = self.model_2_cv1_act(v_5)
        v_7 = self.model_2_m_0_cv1_conv(v_6)
        v_8 = self.model_2_m_0_cv1_act(v_7)
        v_9 = self.model_2_m_0_cv2_conv(v_8)
        v_10 = self.model_2_m_0_cv2_act(v_9)
        v_11 = (v_6 + v_10)
        v_12 = self.model_2_cv2_conv(v_4)
        v_13 = self.model_2_cv2_act(v_12)
        v_14 = torch.cat((v_11, v_13), dim=1)
        v_15 = self.model_2_cv3_conv(v_14)
        v_16 = self.model_2_cv3_act(v_15)
        v_17 = self.model_3_conv(v_16)
        v_18 = self.model_3_act(v_17)
        v_19 = self.model_4_cv1_conv(v_18)
        v_20 = self.model_4_cv1_act(v_19)
        v_21 = self.model_4_m_0_cv1_conv(v_20)
        v_22 = self.model_4_m_0_cv1_act(v_21)
        v_23 = self.model_4_m_0_cv2_conv(v_22)
        v_24 = self.model_4_m_0_cv2_act(v_23)
        v_25 = (v_20 + v_24)
        v_26 = self.model_4_m_1_cv1_conv(v_25)
        v_27 = self.model_4_m_1_cv1_act(v_26)
        v_28 = self.model_4_m_1_cv2_conv(v_27)
        v_29 = self.model_4_m_1_cv2_act(v_28)
        v_30 = (v_25 + v_29)
        v_31 = self.model_4_cv2_conv(v_18)
        v_32 = self.model_4_cv2_act(v_31)
        v_33 = torch.cat((v_30, v_32), dim=1)
        v_34 = self.model_4_cv3_conv(v_33)
        v_35 = self.model_4_cv3_act(v_34)
        v_36 = self.model_5_conv(v_35)
        v_37 = self.model_5_act(v_36)
        v_38 = self.model_6_cv1_conv(v_37)
        v_39 = self.model_6_cv1_act(v_38)
        v_40 = self.model_6_m_0_cv1_conv(v_39)
        v_41 = self.model_6_m_0_cv1_act(v_40)
        v_42 = self.model_6_m_0_cv2_conv(v_41)
        v_43 = self.model_6_m_0_cv2_act(v_42)
        v_44 = (v_39 + v_43)
        v_45 = self.model_6_m_1_cv1_conv(v_44)
        v_46 = self.model_6_m_1_cv1_act(v_45)
        v_47 = self.model_6_m_1_cv2_conv(v_46)
        v_48 = self.model_6_m_1_cv2_act(v_47)
        v_49 = (v_44 + v_48)
        v_50 = self.model_6_m_2_cv1_conv(v_49)
        v_51 = self.model_6_m_2_cv1_act(v_50)
        v_52 = self.model_6_m_2_cv2_conv(v_51)
        v_53 = self.model_6_m_2_cv2_act(v_52)
        v_54 = (v_49 + v_53)
        v_55 = self.model_6_cv2_conv(v_37)
        v_56 = self.model_6_cv2_act(v_55)
        v_57 = torch.cat((v_54, v_56), dim=1)
        v_58 = self.model_6_cv3_conv(v_57)
        v_59 = self.model_6_cv3_act(v_58)
        v_60 = self.model_7_conv(v_59)
        v_61 = self.model_7_act(v_60)
        v_62 = self.model_8_cv1_conv(v_61)
        v_63 = self.model_8_cv1_act(v_62)
        v_64 = self.model_8_m_0_cv1_conv(v_63)
        v_65 = self.model_8_m_0_cv1_act(v_64)
        v_66 = self.model_8_m_0_cv2_conv(v_65)
        v_67 = self.model_8_m_0_cv2_act(v_66)
        v_68 = (v_63 + v_67)
        v_69 = self.model_8_cv2_conv(v_61)
        v_70 = self.model_8_cv2_act(v_69)
        v_71 = torch.cat((v_68, v_70), dim=1)
        v_72 = self.model_8_cv3_conv(v_71)
        v_73 = self.model_8_cv3_act(v_72)
        v_74 = self.model_9_cv1_conv(v_73)
        v_75 = self.model_9_cv1_act(v_74)
        v_76 = self.model_9_m(v_75)
        v_77 = self.pnnx_unique_0(v_76)
        v_78 = self.pnnx_unique_1(v_77)
        v_79 = torch.cat((v_75, v_76, v_77, v_78), dim=1)
        v_80 = self.model_9_cv2_conv(v_79)
        v_81 = self.model_9_cv2_act(v_80)
        v_82 = self.model_10_conv(v_81)
        v_83 = self.model_10_act(v_82)
        v_84 = self.model_11(v_83)
        v_85 = torch.cat((v_84, v_59), dim=1)
        v_86 = self.model_13_cv1_conv(v_85)
        v_87 = self.model_13_cv1_act(v_86)
        v_88 = self.model_13_m_0_cv1_conv(v_87)
        v_89 = self.model_13_m_0_cv1_act(v_88)
        v_90 = self.model_13_m_0_cv2_conv(v_89)
        v_91 = self.model_13_m_0_cv2_act(v_90)
        v_92 = self.model_13_cv2_conv(v_85)
        v_93 = self.model_13_cv2_act(v_92)
        v_94 = torch.cat((v_91, v_93), dim=1)
        v_95 = self.model_13_cv3_conv(v_94)
        v_96 = self.model_13_cv3_act(v_95)
        v_97 = self.model_14_conv(v_96)
        v_98 = self.model_14_act(v_97)
        v_99 = self.model_15(v_98)
        v_100 = torch.cat((v_99, v_35), dim=1)
        v_101 = self.model_17_cv1_conv(v_100)
        v_102 = self.model_17_cv1_act(v_101)
        v_103 = self.model_17_m_0_cv1_conv(v_102)
        v_104 = self.model_17_m_0_cv1_act(v_103)
        v_105 = self.model_17_m_0_cv2_conv(v_104)
        v_106 = self.model_17_m_0_cv2_act(v_105)
        v_107 = self.model_17_cv2_conv(v_100)
        v_108 = self.model_17_cv2_act(v_107)
        v_109 = torch.cat((v_106, v_108), dim=1)
        v_110 = self.model_17_cv3_conv(v_109)
        v_111 = self.model_17_cv3_act(v_110)
        v_112 = self.model_18_conv(v_111)
        v_113 = self.model_18_act(v_112)
        v_114 = torch.cat((v_113, v_98), dim=1)
        v_115 = self.model_20_cv1_conv(v_114)
        v_116 = self.model_20_cv1_act(v_115)
        v_117 = self.model_20_m_0_cv1_conv(v_116)
        v_118 = self.model_20_m_0_cv1_act(v_117)
        v_119 = self.model_20_m_0_cv2_conv(v_118)
        v_120 = self.model_20_m_0_cv2_act(v_119)
        v_121 = self.model_20_cv2_conv(v_114)
        v_122 = self.model_20_cv2_act(v_121)
        v_123 = torch.cat((v_120, v_122), dim=1)
        v_124 = self.model_20_cv3_conv(v_123)
        v_125 = self.model_20_cv3_act(v_124)
        v_126 = self.model_21_conv(v_125)
        v_127 = self.model_21_act(v_126)
        v_128 = torch.cat((v_127, v_83), dim=1)
        v_129 = self.model_23_cv1_conv(v_128)
        v_130 = self.model_23_cv1_act(v_129)
        v_131 = self.model_23_m_0_cv1_conv(v_130)
        v_132 = self.model_23_m_0_cv1_act(v_131)
        v_133 = self.model_23_m_0_cv2_conv(v_132)
        v_134 = self.model_23_m_0_cv2_act(v_133)
        v_135 = self.model_23_cv2_conv(v_128)
        v_136 = self.model_23_cv2_act(v_135)
        v_137 = torch.cat((v_134, v_136), dim=1)
        v_138 = self.model_23_cv3_conv(v_137)
        v_139 = self.model_23_cv3_act(v_138)
        v_140 = self.pnnx_53_data
        v_141 = self.pnnx_54_data
        v_142 = self.pnnx_55_data
        v_143 = self.pnnx_56_data
        v_144 = self.pnnx_58_data
        v_145 = self.pnnx_60_data
        v_146 = self.model_24_m_0(v_111)
        v_147 = v_146.reshape(1, 3, 85, 80, 80)
        v_148 = v_147.permute(dims=(0,1,3,4,2))
        v_149 = F.sigmoid(v_148)
        v_150, v_151, v_152 = torch.split(tensor=v_149, dim=4, split_size_or_sections=(2,2,81))
        v_153 = (((v_150 * 2) + v_145) * 8.0)
        v_154 = (torch.pow((v_151 * 2), 2) * v_144)
        v_155 = torch.cat((v_153, v_154, v_152), dim=4)
        v_156 = v_155.reshape(1, 19200, 85)
        v_157 = self.model_24_m_1(v_125)
        v_158 = v_157.reshape(1, 3, 85, 40, 40)
        v_159 = v_158.permute(dims=(0,1,3,4,2))
        v_160 = F.sigmoid(v_159)
        v_161, v_162, v_163 = torch.split(tensor=v_160, dim=4, split_size_or_sections=(2,2,81))
        v_164 = (((v_161 * 2) + v_143) * 16.0)
        v_165 = (torch.pow((v_162 * 2), 2) * v_142)
        v_166 = torch.cat((v_164, v_165, v_163), dim=4)
        v_167 = v_166.reshape(1, 4800, 85)
        v_168 = self.model_24_m_2(v_139)
        v_169 = v_168.reshape(1, 3, 85, 20, 20)
        v_170 = v_169.permute(dims=(0,1,3,4,2))
        v_171 = F.sigmoid(v_170)
        v_172, v_173, v_174 = torch.split(tensor=v_171, dim=4, split_size_or_sections=(2,2,81))
        v_175 = (((v_172 * 2) + v_141) * 32.0)
        v_176 = (torch.pow((v_173 * 2), 2) * v_140)
        v_177 = torch.cat((v_175, v_176, v_174), dim=4)
        v_178 = v_177.reshape(1, 1200, 85)
        v_179 = torch.cat((v_156, v_167, v_178), dim=1)
        v_180 = (v_179, )
        return v_180

def export_torchscript():
    net = Model()
    net.float()
    net.eval()

    torch.manual_seed(0)
    v_0 = torch.rand(1, 3, 640, 640, dtype=torch.float)

    mod = torch.jit.trace(net, v_0)
    mod.save("yolov5n_pnnx.py.pt")

def export_onnx():
    net = Model()
    net.float()
    net.eval()

    torch.manual_seed(0)
    v_0 = torch.rand(1, 3, 640, 640, dtype=torch.float)

    torch.onnx.export(net, v_0, "yolov5n_pnnx.py.onnx", export_params=True, operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK, opset_version=13, input_names=['in0'], output_names=['out0'])

def export_pnnx():
    net = Model()
    net.float()
    net.eval()

    torch.manual_seed(0)
    v_0 = torch.rand(1, 3, 640, 640, dtype=torch.float)

    import pnnx
    pnnx.export(net, "yolov5n_pnnx.py.pt", v_0)

def export_ncnn():
    export_pnnx()

@torch.no_grad()
def test_inference(self, v_0=None):
        # 如果没有传入输入，默认创建随机数据
        if v_0 is None:
            torch.manual_seed(0)
            v_0 = torch.rand(1, 3, 640, 640, dtype=torch.float)

        return self.forward(v_0)

# 直接运行时，会调用此部分
if __name__ == "__main__":
    net = Model()
    net.eval()
    out = net.test_inference()  # 默认会给出一个随机的输入
    print(out)

显示效果：测试验证成功

bash 复制代码 
(my_pytorch_env) user@user-EliteMini-Series:~/RK/RK3568/4_rk3568_ncnn/rk3568_work$ python compare.py
pnnx shape: torch.Size([1, 25200, 85])
ncnn shape: torch.Size([1, 25200, 85])
Shapes equal: True
Max absolute diff: 0.0010223388671875
Allclose (within 1e-6): True

2.3、创建测试工程

2.3.1、构建目录

复制代码 
rk3568_work
 ├─ ncnn
 ├─ yolov5n.ncnn.param
 ├─ yolov5n.ncnn.bin
 ├─ my_pytorch_env(虚拟环境主要使用pnnx)
 ├─ deploy_ubuntu
 │   ├─ CMakeLists.txt   (等会创建)
 │   ├─ src
 │   │   └─ yolov5_ncnn.cpp (等会创建)
 │   ├─ model
 │   │   ├─ yolov5n.ncnn.param
 │   │   └─ yolov5n.ncnn.bin
 │   └─ test.jpg
 └─ build_ubuntu

2.3.2、构建源码和目录

复制代码 
cd /home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work

# 创建源码目录和构建目录
mkdir -p deploy_ubuntu/src
mkdir -p deploy_ubuntu/model
mkdir -p build_ubuntu

cd /home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/build_ubuntu

cmake ../deploy_ubuntu
cmake --build . -j$(nproc)

效果：出现下列效果是默认你把所有的文件都配置好了，包括图片

bash 复制代码 
user@user-EliteMini-Series:~/RK/RK3568/4_rk3568_ncnn/rk3568_work/build_ubuntu$ ./yolov5_ncnn yolov5n.ncnn.param yolov5n.ncnn.bin bus.jpg 
out shape (c, h, w): 1, 25200, 85
first 10 values: 8.45431 2.88228 4.83573 6.39983 0.9488 2.09542e-09 4.4782e-09 3.80263e-05 2.06879e-06 4.36303e-06

2.3.3、构建CMakeLists.txt

复制代码 
cmake_minimum_required(VERSION 3.16)
project(yolov5n_ncnn_ubuntu CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# ncnn 源码根目录
set(NCNN_ROOT "/home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/ncnn")

# ncnn 在 x86 上的 build 目录
set(NCNN_BUILD "/home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/ncnn-build-host")
set(CMAKE_BUILD_TYPE Release)
add_compile_options(-O3 -DNDEBUG)

#yolov5_ncnn.cpp这个就是类似的main.cpp
add_executable(yolov5_ncnn src/yolov5_ncnn.cpp)


# 这里一定要同时包含 源码/src + build/src
target_include_directories(yolov5_ncnn PRIVATE
    ${NCNN_ROOT}/src
    ${NCNN_BUILD}/src
)

target_link_libraries(yolov5_ncnn PRIVATE
    ${NCNN_BUILD}/src/libncnn.a
    pthread
)

find_package(OpenMP)
if(OpenMP_CXX_FOUND)
    target_link_libraries(yolov5_ncnn PRIVATE OpenMP::OpenMP_CXX)
endif()

find_package(OpenCV REQUIRED)
target_include_directories(yolov5_ncnn PRIVATE ${OpenCV_INCLUDE_DIRS})
target_link_libraries(yolov5_ncnn PRIVATE ${OpenCV_LIBS})

2.3.4、构建build.sh

复制代码 
#!/bin/bash 

cd /home/user/RK/RK3568/4_rk3568_ncnn/rk3568_work/build_ubuntu

cmake ../deploy_ubuntu
cmake --build . -j$(nproc)

#这个表示识别当前这个图片bus.jpg
sudo ./yolov5_ncnn yolov5n.ncnn.param yolov5n.ncnn.bin

2.3.5、构建yolov5_ncnn.cpp

复制代码 
#include <iostream>
#include <vector>
#include <string>
#include <algorithm>
#include <cmath>
#include <cstdio>

#include <opencv2/opencv.hpp>
#include "net.h"    // ncnn 头文件

// ------------------------ COCO 80 类名字 ------------------------
static const std::vector<std::string> coco_class_names = {
    "person", "bicycle", "car", "motorbike", "aeroplane", "bus", "train",
    "truck", "boat", "traffic light", "fire hydrant", "stop sign",
    "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep",
    "cow", "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella",
    "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
    "sports ball", "kite", "baseball bat", "baseball glove", "skateboard",
    "surfboard", "tennis racket", "bottle", "wine glass", "cup", "fork",
    "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange",
    "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair",
    "sofa", "pottedplant", "bed", "diningtable", "toilet", "tvmonitor",
    "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave",
    "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase",
    "scissors", "teddy bear", "hair drier", "toothbrush"
};

// ------------------------ 基本结构体和函数 ------------------------

// 检测框结构体：左上角(x1,y1) 和 右下角(x2,y2)，在原图坐标系下
struct Detection {
    float x1, y1, x2, y2;
    int   class_id;
    float score;        // obj_conf * class_conf
};

// 计算 IOU
float intersection_over_union(const Detection& a, const Detection& b)
{
    float x1 = std::max(a.x1, b.x1);
    float y1 = std::max(a.y1, b.y1);
    float x2 = std::min(a.x2, b.x2);
    float y2 = std::min(a.y2, b.y2);

    float w = std::max(0.f, x2 - x1);
    float h = std::max(0.f, y2 - y1);
    float inter = w * h;
    float area_a = (a.x2 - a.x1) * (a.y2 - a.y1);
    float area_b = (b.x2 - b.x1) * (b.y2 - b.y1);
    float uni = area_a + area_b - inter;

    if (uni <= 0.f) return 0.f;
    return inter / uni;
}

/**
 * 解码 YOLOv5 输出 + NMS
 *  out:       ncnn::Mat，shape 为 (h=num_preds=25200, w=num_attr=85, c=1)
 *  detections: 输出的最终检测结果
 *  conf_thresh: 置信度阈值（与 Python detect.py 对齐，0.25）
 *  nms_thresh:  NMS IOU 阈值（默认 0.45）
 *  orig_w, orig_h: 原图宽高
 *  scale:     letterbox 时的缩放比例 r
 *  pad_x, pad_y: letterbox 左右 / 上下 padding 的像素（dx, dy）
 *  input_size:  网络输入尺寸（通常 640）
 */
void decode_and_nms(const ncnn::Mat& out,
                    std::vector<Detection>& detections,
                    float conf_thresh,
                    float nms_thresh,
                    int orig_w,
                    int orig_h,
                    float scale,
                    int pad_x,
                    int pad_y,
                    int input_size = 640)
{
    // 这里根据你打印的结果：out.w = 85, out.h = 25200, out.c = 1
    int num_preds = out.h;   // 25200
    int num_attr  = out.w;   // 85
    const int num_classes = num_attr - 5;   // 80

    for (int i = 0; i < num_preds; i++)
    {
        const float* ptr = out.row(i);

        float obj_conf = ptr[4];

        // 和 Python 行为更接近：不再用 0.01 的硬编码阈值，而是和 conf_thresh 一致
        if (obj_conf < conf_thresh)
            continue;

        // 找类别概率最大的那个类
        int   class_id     = -1;
        float max_cls_conf = 0.f;
        for (int c = 0; c < num_classes; c++)
        {
            float cls_conf = ptr[5 + c];
            if (cls_conf > max_cls_conf)
            {
                max_cls_conf = cls_conf;
                class_id = c;
            }
        }

        // final score = obj_conf * class_conf
        float score = obj_conf * max_cls_conf;
        if (score < conf_thresh)
            continue;

        // YOLOv5 导出到 ncnn 时，前 4 个通常是 cx, cy, w, h（在 640x640 上的像素坐标）
        float cx = ptr[0];
        float cy = ptr[1];
        float w  = ptr[2];
        float h  = ptr[3];

        // 在 640x640 上的左上/右下
        float x1 = cx - w * 0.5f;
        float y1 = cy - h * 0.5f;
        float x2 = cx + w * 0.5f;
        float y2 = cy + h * 0.5f;

        // 去掉 letterbox padding，再除以 scale，还原到原图坐标
        x1 = (x1 - pad_x) / scale;
        y1 = (y1 - pad_y) / scale;
        x2 = (x2 - pad_x) / scale;
        y2 = (y2 - pad_y) / scale;

        // 裁剪到原图范围
        x1 = std::max(0.f, std::min(x1, (float)orig_w - 1));
        y1 = std::max(0.f, std::min(y1, (float)orig_h - 1));
        x2 = std::max(0.f, std::min(x2, (float)orig_w - 1));
        y2 = std::max(0.f, std::min(y2, (float)orig_h - 1));

        // 防止异常框
        if (x2 <= x1 || y2 <= y1)
            continue;

        Detection det;
        det.x1 = x1;
        det.y1 = y1;
        det.x2 = x2;
        det.y2 = y2;
        det.class_id = class_id;
        det.score = score;
        detections.push_back(det);
    }

    // 按 score 降序排序
    std::sort(detections.begin(), detections.end(),
              [](const Detection& a, const Detection& b) {
                  return a.score > b.score;
              });

    // NMS
    std::vector<Detection> nms_result;
    for (const auto& det : detections)
    {
        bool keep = true;
        for (const auto& kept : nms_result)
        {
            float iou = intersection_over_union(det, kept);
            if (iou > nms_thresh)
            {
                keep = false;
                break;
            }
        }
        if (keep)
            nms_result.push_back(det);
    }

    detections.swap(nms_result);
}

// ------------------------ draw_detections_and_print ------------------------

// 为每个类别分配一个唯一的颜色
cv::Scalar get_class_color(int class_id)
{
    // 颜色池，根据 class_id 获取不同的颜色
    std::vector<cv::Scalar> colors = {
        cv::Scalar(0, 0, 255),   // Red    -> person
        cv::Scalar(0, 255, 0),   // Green  -> bicycle
        cv::Scalar(255, 0, 0),   // Blue   -> car
        cv::Scalar(0, 255, 255), // Yellow -> motorbike
        cv::Scalar(255, 255, 0), // Cyan   -> aeroplane
        cv::Scalar(255, 0, 255), // Magenta -> bus
        // 其他颜色可以继续添加
    };
    
    // 保证类别 ID 不超出颜色池
    return colors[class_id % colors.size()];
}



// 把检测结果画到图片上，并在控制台打印
void draw_detections_and_print(cv::Mat& img,
                               const std::vector<Detection>& dets)
{
    for (const auto& det : dets)
    {
        int x1 = std::max(0, (int)std::round(det.x1));
        int y1 = std::max(0, (int)std::round(det.y1));
        int x2 = std::min(img.cols - 1, (int)std::round(det.x2));
        int y2 = std::min(img.rows - 1, (int)std::round(det.y2));

        // 获取每个类别对应的颜色
        cv::Scalar color = get_class_color(det.class_id);

        // 画矩形框
        cv::rectangle(img,
                      cv::Point(x1, y1),
                      cv::Point(x2, y2),
                      color, 2);  // 使用类别颜色绘制框

        // 类别名字
        std::string cls_name = "cls_" + std::to_string(det.class_id);
        if (det.class_id >= 0 &&
            det.class_id < (int)coco_class_names.size())
        {
            cls_name = coco_class_names[det.class_id];
        }

        // 文本：类名 + 分数
        char text[256];
        std::snprintf(text, sizeof(text), "%s %.2f",
                      cls_name.c_str(), det.score);

        int baseline = 0;
        cv::Size tsize = cv::getTextSize(text,
                                         cv::FONT_HERSHEY_SIMPLEX,
                                         0.5,
                                         1,
                                         &baseline);
        int ty = std::max(0, y1 - tsize.height - baseline);

        // 文本背景
        cv::rectangle(img,
                      cv::Point(x1, ty),
                      cv::Point(x1 + tsize.width,
                                ty + tsize.height + baseline),
                      color, -1);  // 使用类别颜色作为文本背景

        // 文字
        cv::putText(img, text,
                    cv::Point(x1, ty + tsize.height),
                    cv::FONT_HERSHEY_SIMPLEX,
                    0.5,
                    cv::Scalar(0, 0, 0),
                    1);

        // 控制台打印
        std::printf("detect: class=%s (id=%d) score=%.3f box=[%d,%d,%d,%d]\n",
                    cls_name.c_str(), det.class_id, det.score,
                    x1, y1, x2, y2);
    }
}

// ------------------------ main ------------------------

int main(int argc, char** argv)
{
    if (argc < 4)
    {
        std::cerr << "Usage: " << argv[0]
                  << " yolov5n.ncnn.param yolov5n.ncnn.bin image.jpg"
                  << std::endl;
        return -1;
    }

    const char* param_path = argv[1];
    const char* bin_path   = argv[2];
    const char* image_path = argv[3];

    // 1. 读入原图（BGR）
    cv::Mat img_bgr = cv::imread(image_path);
    if (img_bgr.empty())
    {
        std::cerr << "Error: unable to read image: "
                  << image_path << std::endl;
        return -1;
    }

    int orig_w = img_bgr.cols;
    int orig_h = img_bgr.rows;

    // 2. 初始化 ncnn 网络
    ncnn::Net net;
    if (net.load_param(param_path) != 0)
    {
        std::cerr << "Error: failed to load ncnn param." << std::endl;
        return -1;
    }
    if (net.load_model(bin_path) != 0)
    {
        std::cerr << "Error: failed to load ncnn bin." << std::endl;
        return -1;
    }

    // 3. 前处理：letterbox 到 640×640（和 Python 对齐）
    const int input_size = 640;

    // 计算等比缩放比例
    float r = std::min((float)input_size / (float)orig_w,
                       (float)input_size / (float)orig_h);
    int new_w = (int)std::round(orig_w * r);
    int new_h = (int)std::round(orig_h * r);

    cv::Mat resized_bgr;
    cv::resize(img_bgr, resized_bgr, cv::Size(new_w, new_h));

    // 生成 640x640 的画布（用 114 填充，和 YOLOv5 一致）
    cv::Mat letterbox_bgr(input_size, input_size,
                          CV_8UC3,
                          cv::Scalar(114, 114, 114));

    int dx = (input_size - new_w) / 2;
    int dy = (input_size - new_h) / 2;

    resized_bgr.copyTo(letterbox_bgr(cv::Rect(dx, dy, new_w, new_h)));

    // ⭐ 与 Python 对齐的关键改动：BGR -> RGB
    cv::Mat letterbox_rgb;
    cv::cvtColor(letterbox_bgr, letterbox_rgb, cv::COLOR_BGR2RGB);

    // 转成 ncnn::Mat，并做 /255 归一化
    ncnn::Mat in = ncnn::Mat::from_pixels(
        letterbox_rgb.data,
        ncnn::Mat::PIXEL_RGB,   // ⭐ 改为 RGB
        input_size,
        input_size);

    const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};
    in.substract_mean_normalize(nullptr, norm_vals);

    // 4. 推理
    ncnn::Extractor ex = net.create_extractor();
    int ret = ex.input("in0", in);   // 注意：与 .param 输入 blob 名一致
    if (ret != 0)
    {
        std::cerr << "ex.input failed, ret = " << ret << std::endl;
        return -1;
    }

    ncnn::Mat out;
    ret = ex.extract("out0", out);   // 注意：与 .param 输出 blob 名一致
    std::cout << "extract ret = " << ret
              << "  out.w = " << out.w
              << "  out.h = " << out.h
              << "  out.c = " << out.c << std::endl;
    if (ret != 0 || out.w == 0 || out.h == 0)
    {
        std::cerr << "Error: extract output failed." << std::endl;
        return -1;
    }

    // 5. 解码 + NMS
    std::vector<Detection> detections;
    float conf_thresh = 0.25f;   // 与 Python detect.py 一致
    float nms_thresh  = 0.45f;
    decode_and_nms(out, detections,
                   conf_thresh, nms_thresh,
                   orig_w, orig_h,
                   r, dx, dy,
                   input_size);

    std::cout << "total detections after NMS: "
              << detections.size() << std::endl;

    // 6. 画框 + 打印（在 BGR 原图上画）
    cv::Mat img_show = img_bgr.clone();
    draw_detections_and_print(img_show, detections);

    // 7. 保存结果图
    std::string out_path = "output.jpg";
    cv::imwrite(out_path, img_show);
    std::cout << "Detection complete, output saved as "
              << out_path << std::endl;

    return 0;
}

2.3.6、使用上述步骤，在ubuntu上的效果

识别图片效果：能识别到，说明推理正常

换一张也能识别，虽然识别的不是非常准确（这个githup获取的yolov5n.pt，没有手动训练过，出现不精准，属于正常）

2.4、解释输出现象（YOLOv5 head 的原始输出（已经 decode+sigmoid+scale）但未经过 NMS）

bash 复制代码 
user@user-EliteMini-Series:~/RK/RK3568/4_rk3568_ncnn/rk3568_work/build_ubuntu$ ./yolov5_ncnn yolov5n.ncnn.param yolov5n.ncnn.bin bus.jpg 
out shape (c, h, w): 1, 25200, 85
first 10 values: 8.45431 2.88228 4.83573 6.39983 0.9488 2.09542e-09 4.4782e-09 3.80263e-05 2.06879e-06 4.36303e-06

2.4.1、输出形状：out shape (c, h, w): 1, 25200, 85

维度 数值 解释
c 1 batch 或通道（ncnn 的内部格式）
h 25200 预测点数量（即 80x80 + 40x40 + 20x20 共 3 个输出层合并后的 anchor-free 点）
w 85 每个预测点的属性数量（xywh + obj + 80 类 = 85）

2.4.2、为什么是 25200？

YOLOv5n 有三个检测输出层：

特征图大小 锚点类型 网格点数量
80×80 P3 6400
40×40 P4 1600
20×20 P5 400

每层 3 个 anchor → 等价拆成单 anchor，再展平成：

复制代码 
80×80×3 = 19200
40×40×3 = 4800
20×20×3 = 1200
总计：19200 + 4800 +1200 = 25200

2.4.3、为什么每个点是 85 维？

YOLOv5 的每个预测值的结构：

复制代码 
[ x, y, w, h, obj_conf, class_prob_0, class_prob_1, ... class_prob_79 ]
[ 0  1  2  3  4         5             第1个类别         第79个类别 ]
= 4 + 1 + 80 = 85

你打印的前 10 个值对应：

索引 数值 含义
0 8.45431 x 坐标（激活后）
1 2.88228 y 坐标（激活后）
2 4.83573 width（激活后）
3 6.39983 height（激活后）
4 0.9488 objectness 置信度
5 2.09542e-09 第 0 类概率
6 4.4782e-09 第 1 类概率
7 3.80263e-05 第 2 类概率
8 2.06879e-06 第 3 类概率
9 4.36303e-06 第 4 类概率

2.4.4、这些值的类型是什么？

复制代码 
1. 所有数值类型都来自 **`ncnn::Mat`**，底层是：float（32位浮点数）
2. ncnn 的 Mat 是 CHW 格式：c=1, h=25200, w=85，第 0 通道下的所有 (h,w) 数字（也就是[25200][85]的二维数组）
3. 第 5 个值 0.9488 是 objectness 分数，第 6~85 个值是 80 个类别的置信度（还未乘以 objectness）
4.
索引 含义
0 预测框中心 x
1 预测框中心 y
2 预测框宽度
3 高度
4 objectness 置信度
5~84 80 类概率

类型全部为 float32。

相关推荐
Coding茶水间6 小时前
基于深度学习的遥感地面物体检测系统演示与介绍(YOLOv12/v11/v8/v5模型+Pyqt5界面+训练代码+数据集)
图像处理·人工智能·深度学习·yolo·目标检测·计算机视觉
咨询QQ:4877392786 小时前
基于FPGA的FOC电流环实现：Verilog编写、SVPWM算法、ADC采样、串口通信、Si...
yolo
前网易架构师-高司机8 小时前
水下异物识别数据集，识别率83.4%可识别口罩,手机,瓶,手套,金属,网 袋,塑料,杆,太阳镜,轮胎等常见异物并分类，支持yolo，json，xml格式的标注
yolo·数据集·垃圾·水里·异物·杂物·水下
深度学习lover9 小时前
＜数据集＞yolo茶叶嫩芽识别数据集＜目标检测＞
人工智能·python·yolo·目标检测·计算机视觉·茶叶嫩芽识别
a1111111111ss10 小时前
SCSAyolo
yolo
胖墩会武术11 小时前
【PyTorch项目实战】Ultralytics YOLO模型本地部署（训练 + 测试 + 推理）
人工智能·pytorch·yolo
magic_ll12 小时前
【论文阅读】【yolo系列】YOLOv10: Real-Time End-to-End Object Detection
论文阅读·yolo·目标检测
AI小怪兽13 小时前
基于YOLO的小目标检测增强：一种提升精度与效率的新框架
人工智能·深度学习·yolo·目标检测·计算机视觉
热门推荐
01GitHub 镜像站点02【超详细教程】手把手教你从微软官网免费下载Windows 10官方原版ISO镜像（2025最新版）03UV安装并设置国内源04【AutoGLM部署】本地私有化部署AI手机Agent05安娜的档案(Anna’s Archive) 镜像网站/国内最新可访问入口（持续更新）06Linux下V2Ray安装配置指南07React CVE-2025-55182漏洞排查与修复指南08BongoCat - 跨平台键盘猫动画工具09Open-AutoGLM Windows 安装部署教程10在VSCode配置Java开发环境的保姆级教程（适配各类AI编程IDE）